“mcs” — 2017/6/5 — 19:42 — page i — #1
Mathematics for Computer Science
revised Monday 5th June, 2017, 19:42
Eric Lehman
Google Inc.
F Thomson Leighton
Department of Mathematics
and the Computer Science and AI Laboratory,
Massachussetts Institute of Technology;
Akamai Technologies
Albert R Meyer
Department of Electrical Engineering and Computer Science
and the Computer Science and AI Laboratory,
Massachussetts Institute of Technology
2017, Eric Lehman, F Tom Leighton, Albert R Meyer. This work is available under the
terms of the Creative Commons Attribution-ShareAlike 3.0 license.
“mcs” — 2017/6/5 — 19:42 — page ii — #2
“mcs” — 2017/6/5 — 19:42 — page iii — #3
Contents
I
Proofs
Introduction 3
0.1
1
Well Ordering Proofs
29
Template for WOP Proofs
30
Factoring into Primes
32
Well Ordered Sets
33
Logical Formulas 47
3.1
3.2
3.3
3.4
3.5
3.6
3.7
4
Propositions
5
Predicates
8
The Axiomatic Method
8
Our Axioms
9
Proving an Implication
11
Proving an “If and Only If”
13
Proof by Cases
15
Proof by Contradiction
16
Good Proofs in Practice
17
References
19
The Well Ordering Principle 29
2.1
2.2
2.3
2.4
3
4
What is a Proof? 5
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
1.10
2
References
Propositions from Propositions
48
Propositional Logic in Computer Programs
Equivalence and Validity
54
The Algebra of Propositions
57
The SAT Problem
62
Predicate Formulas
63
References
68
Mathematical Data Types 97
4.1
4.2
4.3
4.4
4.5
Sets
97
Sequences 102
Functions 103
Binary Relations 105
Finite Cardinality 109
52
“mcs” — 2017/6/5 — 19:42 — page iv — #4
iv
Contents
5
Induction 131
5.1
5.2
5.3
6
States and Transitions 167
The Invariant Principle 168
Partial Correctness & Termination 176
The Stable Marriage Problem 181
Recursive Data Types 211
7.1
7.2
7.3
7.4
7.5
7.6
8
Recursive Definitions and Structural Induction 211
Strings of Matched Brackets 215
Recursive Functions on Nonnegative Integers 219
Arithmetic Expressions 221
Games as a Recursive Data Type 226
Induction in Computer Science 230
Infinite Sets 257
8.1
8.2
8.3
8.4
Infinite Cardinality 258
The Halting Problem 267
The Logic of Sets 271
Does All This Really Work?
275
II Structures
Introduction 299
9
147
State Machines 167
6.1
6.2
6.3
6.4
7
Ordinary Induction 131
Strong Induction 140
Strong Induction vs. Induction vs. Well Ordering
Number Theory 301
9.1
9.2
9.3
9.4
9.5
9.6
9.7
9.8
9.9
9.10
9.11
Divisibility 301
The Greatest Common Divisor 306
Prime Mysteries 313
The Fundamental Theorem of Arithmetic 315
Alan Turing 318
Modular Arithmetic 322
Remainder Arithmetic 324
Turing’s Code (Version 2.0) 327
Multiplicative Inverses and Cancelling 329
Euler’s Theorem 333
RSA Public Key Encryption 338
“mcs” — 2017/6/5 — 19:42 — page v — #5
v
Contents
9.12 What has SAT got to do with it?
9.13 References 341
340
10 Directed graphs & Partial Orders 381
10.1 Vertex Degrees 383
10.2 Walks and Paths 384
10.3 Adjacency Matrices 387
10.4 Walk Relations 390
10.5 Directed Acyclic Graphs & Scheduling 391
10.6 Partial Orders 399
10.7 Representing Partial Orders by Set Containment
10.8 Linear Orders 404
10.9 Product Orders 404
10.10 Equivalence Relations 405
10.11 Summary of Relational Properties 407
10.12 References 409
403
11 Communication Networks 441
11.1 Routing 441
11.2 Routing Measures 442
11.3 Network Designs 445
12 Simple Graphs 461
12.1 Vertex Adjacency and Degrees 461
12.2 Sexual Demographics in America 463
12.3 Some Common Graphs 465
12.4 Isomorphism 466
12.5 Bipartite Graphs & Matchings 469
12.6 Coloring 474
12.7 Walks in Simple Graphs 478
12.8 Connectivity 480
12.9 Special Walks and Tours 483
12.10 k-connected Graphs 485
12.11 Forests & Trees 487
12.12 References 494
13 Planar Graphs 533
13.1
13.2
13.3
13.4
13.5
Drawing Graphs in the Plane 533
Definitions of Planar Graphs 533
Euler’s Formula 544
Bounding the Number of Edges in a Planar Graph
Returning to K5 and K3;3 546
545
“mcs” — 2017/6/5 — 19:42 — page vi — #6
vi
Contents
13.6 Coloring Planar Graphs 547
13.7 Classifying Polyhedra 549
13.8 Another Characterization for Planar Graphs
552
III Counting
Introduction 561
14 Sums and Asymptotics 563
14.1
14.2
14.3
14.4
14.5
14.6
14.7
The Value of an Annuity 564
Sums of Powers 570
Approximating Sums 572
Hanging Out Over the Edge 576
Products 582
Double Trouble 585
Asymptotic Notation 588
15 Cardinality Rules 613
15.1 Counting One Thing by Counting Another
15.2 Counting Sequences 614
15.3 The Generalized Product Rule 617
15.4 The Division Rule 621
15.5 Counting Subsets 623
15.6 Sequences with Repetitions 625
15.7 Counting Practice: Poker Hands 629
15.8 The Pigeonhole Principle 634
15.9 Inclusion-Exclusion 643
15.10 Combinatorial Proofs 649
15.11 References 653
16 Generating Functions 693
16.1
16.2
16.3
16.4
16.5
16.6
Infinite Series 693
Counting with Generating Functions
Partial Fractions 701
Solving Linear Recurrences 704
Formal Power Series 709
References 712
695
613
“mcs” — 2017/6/5 — 19:42 — page vii — #7
vii
Contents
IV Probability
Introduction 729
17 Events and Probability Spaces 731
17.1
17.2
17.3
17.4
17.5
17.6
Let’s Make a Deal 731
The Four Step Method 732
Strange Dice 741
The Birthday Principle 748
Set Theory and Probability 750
References 754
18 Conditional Probability 763
18.1
18.2
18.3
18.4
18.5
18.6
18.7
18.8
18.9
Monty Hall Confusion 763
Definition and Notation 764
The Four-Step Method for Conditional Probability
Why Tree Diagrams Work 768
The Law of Total Probability 776
Simpson’s Paradox 778
Independence 780
Mutual Independence 782
Probability versus Confidence 786
19 Random Variables 815
19.1
19.2
19.3
19.4
19.5
Random Variable Examples 815
Independence 817
Distribution Functions 819
Great Expectations 827
Linearity of Expectation 839
20 Deviation from the Mean 871
20.1
20.2
20.3
20.4
20.5
20.6
20.7
Markov’s Theorem 871
Chebyshev’s Theorem 874
Properties of Variance 878
Estimation by Random Sampling 884
Confidence in an Estimation 888
Sums of Random Variables 889
Really Great Expectations 899
21 Random Walks 923
21.1 Gambler’s Ruin
923
766
“mcs” — 2017/6/5 — 19:42 — page viii — #8
viii
Contents
21.2 Random Walks on Graphs
933
V Recurrences
Introduction 951
22 Recurrences 953
22.1
22.2
22.3
22.4
22.5
The Towers of Hanoi 953
Merge Sort 956
Linear Recurrences 960
Divide-and-Conquer Recurrences
A Feel for Recurrences 974
Bibliography 981
Glossary of Symbols 985
Index 989
967
“mcs” — 2017/6/5 — 19:42 — page 1 — #9
I
Proofs
“mcs” — 2017/6/5 — 19:42 — page 2 — #10
“mcs” — 2017/6/5 — 19:42 — page 3 — #11
Introduction
This text explains how to use mathematical models and methods to analyze problems that arise in computer science. Proofs play a central role in this work because
the authors share a belief with most mathematicians that proofs are essential for
genuine understanding. Proofs also play a growing role in computer science; they
are used to certify that software and hardware will always behave correctly, something that no amount of testing can do.
Simply put, a proof is a method of establishing truth. Like beauty, “truth” sometimes depends on the eye of the beholder, and it should not be surprising that what
constitutes a proof differs among fields. For example, in the judicial system, legal
truth is decided by a jury based on the allowable evidence presented at trial. In the
business world, authoritative truth is specified by a trusted person or organization,
or maybe just your boss. In fields such as physics or biology, scientific truth is
confirmed by experiment.1 In statistics, probable truth is established by statistical
analysis of sample data.
Philosophical proof involves careful exposition and persuasion typically based
on a series of small, plausible arguments. The best example begins with “Cogito
ergo sum,” a Latin sentence that translates as “I think, therefore I am.” This phrase
comes from the beginning of a 17th century essay by the mathematician/philosopher,
René Descartes, and it is one of the most famous quotes in the world: do a web
search for it, and you will be flooded with hits.
Deducing your existence from the fact that you’re thinking about your existence
is a pretty cool and persuasive-sounding idea. However, with just a few more lines
1 Actually, only scientific falsehood
can be demonstrated by an experiment—when the experiment
fails to behave as predicted. But no amount of experiment can confirm that the next experiment won’t
fail. For this reason, scientists rarely speak of truth, but rather of theories that accurately predict past,
and anticipated future, experiments.
“mcs” — 2017/6/5 — 19:42 — page 4 — #12
4
0.1. References
of argument in this vein, Descartes goes on to conclude that there is an infinitely
beneficent God. Whether or not you believe in an infinitely beneficent God, you’ll
probably agree that any very short “proof” of God’s infinite beneficence is bound
to be far-fetched. So even in masterful hands, this approach is not reliable.
Mathematics has its own specific notion of “proof.”
Definition. A mathematical proof of a proposition is a chain of logical deductions
leading to the proposition from a base set of axioms.
The three key ideas in this definition are highlighted: proposition, logical deduction, and axiom. Chapter 1 examines these three ideas along with some basic ways
of organizing proofs. Chapter 2 introduces the Well Ordering Principle, a basic
method of proof; later, Chapter 5 introduces the closely related proof method of
induction.
If you’re going to prove a proposition, you’d better have a precise understanding of what the proposition means. To avoid ambiguity and uncertain definitions
in ordinary language, mathematicians use language very precisely, and they often
express propositions using logical formulas; these are the subject of Chapter 3.
The first three Chapters assume the reader is familiar with a few mathematical
concepts like sets and functions. Chapters 4 and 8 offer a more careful look at
such mathematical data types, examining in particular properties and methods for
proving things about infinite sets. Chapter 7 goes on to examine recursively defined
data types.
0.1
References
[14], [49], [1]
“mcs” — 2017/6/5 — 19:42 — page 5 — #13
1
1.1
What is a Proof?
Propositions
Definition. A proposition is a statement (communication) that is either true or
false.
For example, both of the following statements are propositions. The first is true,
and the second is false.
Proposition 1.1.1. 2 + 3 = 5.
Proposition 1.1.2. 1 + 1 = 3.
Being true or false doesn’t sound like much of a limitation, but it does exclude
statements such as “Wherefore art thou Romeo?” and “Give me an A!” It also excludes statements whose truth varies with circumstance such as, “It’s five o’clock,”
or “the stock market will rise tomorrow.”
Unfortunately it is not always easy to decide if a claimed proposition is true or
false:
Claim 1.1.3. For every nonnegative integer n the value of n2 C n C 41 is prime.
(A prime is an integer greater than 1 that is not divisible by any other integer
greater than 1. For example, 2, 3, 5, 7, 11, are the first five primes.) Let’s try some
numerical experimentation to check this proposition. Let
p.n/ WWD n2 C n C 41:1
(1.1)
We begin with p.0/ D 41, which is prime; then
p.1/ D 43; p.2/ D 47; p.3/ D 53; : : : ; p.20/ D 461
are each prime. Hmmm, starts to look like a plausible claim. In fact we can keep
checking through n D 39 and confirm that p.39/ D 1601 is prime.
But p.40/ D 402 C 40 C 41 D 41 41, which is not prime. So Claim 1.1.3
is false since it’s not true that p.n/ is prime for all nonnegative integers n. In
fact, it’s not hard to show that no polynomial with integer coefficients can map all
1 The
symbol WWD means “equal by definition.” It’s always ok simply to write “=” instead of WWD,
but reminding the reader that an equality holds by definition can be helpful.
“mcs” — 2017/6/5 — 19:42 — page 6 — #14
6
Chapter 1
What is a Proof?
nonnegative numbers into prime numbers, unless it’s a constant (see Problem 1.26).
But this example highlights the point that, in general, you can’t check a claim about
an infinite set by checking a finite sample of its elements, no matter how large the
sample.
By the way, propositions like this about all numbers or all items of some kind are
so common that there is a special notation for them. With this notation, Claim 1.1.3
would be
8n 2 N: p.n/ is prime:
(1.2)
Here the symbol 8 is read “for all.” The symbol N stands for the set of nonnegative
integers: 0, 1, 2, 3, . . . (ask your instructor for the complete list). The symbol “2”
is read as “is a member of,” or “belongs to,” or simply as “is in.” The period after
the N is just a separator between phrases.
Here are two even more extreme examples:
Conjecture. [Euler] The equation
a4 C b 4 C c 4 D d 4
has no solution when a; b; c; d are positive integers.
Euler (pronounced “oiler”) conjectured this in 1769. But the conjecture was
proved false 218 years later by Noam Elkies at a liberal arts school up Mass Ave.
The solution he found was a D 95800; b D 217519; c D 414560; d D 422481.
In logical notation, Euler’s Conjecture could be written,
8a 2 ZC 8b 2 ZC 8c 2 ZC 8d 2 ZC : a4 C b 4 C c 4 ¤ d 4 :
Here, ZC is a symbol for the positive integers. Strings of 8’s like this are usually
abbreviated for easier reading:
8a; b; c; d 2 ZC : a4 C b 4 C c 4 ¤ d 4 :
Here’s another claim which would be hard to falsify by sampling: the smallest
possible x; y; z that satisfy the equality each have more than 1000 digits!
False Claim. 313.x 3 C y 3 / D z 3 has no solution when x; y; z 2 ZC .
It’s worth mentioning a couple of further famous propositions whose proofs were
sought for centuries before finally being discovered:
Proposition 1.1.4 (Four Color Theorem). Every map can be colored with 4 colors
so that adjacent2 regions have different colors.
2 Two
regions are adjacent only when they share a boundary segment of positive length. They are
not considered to be adjacent if their boundaries meet only at a few points.
“mcs” — 2017/6/5 — 19:42 — page 7 — #15
1.1. Propositions
7
Several incorrect proofs of this theorem have been published, including one that
stood for 10 years in the late 19th century before its mistake was found. A laborious
proof was finally found in 1976 by mathematicians Appel and Haken, who used a
complex computer program to categorize the four-colorable maps. The program
left a few thousand maps uncategorized, which were checked by hand by Haken
and his assistants—among them his 15-year-old daughter.
There was reason to doubt whether this was a legitimate proof—the proof was
too big to be checked without a computer. No one could guarantee that the computer calculated correctly, nor was anyone enthusiastic about exerting the effort
to recheck the four-colorings of thousands of maps that were done by hand. Two
decades later a mostly intelligible proof of the Four Color Theorem was found,
though a computer is still needed to check four-colorability of several hundred special maps.3
Proposition 1.1.5 (Fermat’s Last Theorem). There are no positive integers x, y
and z such that
xn C yn D zn
for some integer n > 2.
In a book he was reading around 1630, Fermat claimed to have a proof for this
proposition, but not enough space in the margin to write it down. Over the years,
the Theorem was proved to hold for all n up to 4,000,000, but we’ve seen that this
shouldn’t necessarily inspire confidence that it holds for all n. There is, after all,
a clear resemblance between Fermat’s Last Theorem and Euler’s false Conjecture.
Finally, in 1994, British mathematician Andrew Wiles gave a proof, after seven
years of working in secrecy and isolation in his attic. His proof did not fit in any
margin.4
Finally, let’s mention another simply stated proposition whose truth remains unknown.
Conjecture 1.1.6 (Goldbach). Every even integer greater than 2 is the sum of two
primes.
Goldbach’s Conjecture dates back to 1742. It is known to hold for all numbers
up to 1018 , but to this day, no one knows whether it’s true or false.
3 The story of the proof of the Four Color Theorem is told in a well-reviewed popular (nontechnical) book: “Four Colors Suffice. How the Map Problem was Solved.” Robin Wilson. Princeton
Univ. Press, 2003, 276pp. ISBN 0-691-11533-8.
4 In fact, Wiles’ original proof was wrong, but he and several collaborators used his ideas to arrive
at a correct proof a year later. This story is the subject of the popular book, Fermat’s Enigma by
Simon Singh, Walker & Company, November, 1997.
“mcs” — 2017/6/5 — 19:42 — page 8 — #16
8
Chapter 1
What is a Proof?
For a computer scientist, some of the most important things to prove are the
correctness of programs and systems—whether a program or system does what it’s
supposed to. Programs are notoriously buggy, and there’s a growing community
of researchers and practitioners trying to find ways to prove program correctness.
These efforts have been successful enough in the case of CPU chips that they are
now routinely used by leading chip manufacturers to prove chip correctness and
avoid some notorious past mistakes.
Developing mathematical methods to verify programs and systems remains an
active research area. We’ll illustrate some of these methods in Chapter 5.
1.2
Predicates
A predicate can be understood as a proposition whose truth depends on the value
of one or more variables. So “n is a perfect square” describes a predicate, since you
can’t say if it’s true or false until you know what the value of the variable n happens
to be. Once you know, for example, that n equals 4, the predicate becomes the true
proposition “4 is a perfect square”. Remember, nothing says that the proposition
has to be true: if the value of n were 5, you would get the false proposition “5 is a
perfect square.”
Like other propositions, predicates are often named with a letter. Furthermore, a
function-like notation is used to denote a predicate supplied with specific variable
values. For example, we might use the name “P ” for predicate above:
P .n/ WWD “n is a perfect square”;
and repeat the remarks above by asserting that P .4/ is true, and P .5/ is false.
This notation for predicates is confusingly similar to ordinary function notation.
If P is a predicate, then P .n/ is either true or false, depending on the value of n.
On the other hand, if p is an ordinary function, like n2 C1, then p.n/ is a numerical
quantity. Don’t confuse these two!
1.3
The Axiomatic Method
The standard procedure for establishing truth in mathematics was invented by Euclid, a mathematician working in Alexandria, Egypt around 300 BC. His idea was
to begin with five assumptions about geometry, which seemed undeniable based on
direct experience. (For example, “There is a straight line segment between every
“mcs” — 2017/6/5 — 19:42 — page 9 — #17
1.4. Our Axioms
9
pair of points”.) Propositions like these that are simply accepted as true are called
axioms.
Starting from these axioms, Euclid established the truth of many additional propositions by providing “proofs.” A proof is a sequence of logical deductions from
axioms and previously proved statements that concludes with the proposition in
question. You probably wrote many proofs in high school geometry class, and
you’ll see a lot more in this text.
There are several common terms for a proposition that has been proved. The
different terms hint at the role of the proposition within a larger body of work.
Important true propositions are called theorems.
A lemma is a preliminary proposition useful for proving later propositions.
A corollary is a proposition that follows in just a few logical steps from a
theorem.
These definitions are not precise. In fact, sometimes a good lemma turns out to be
far more important than the theorem it was originally used to prove.
Euclid’s axiom-and-proof approach, now called the axiomatic method, remains
the foundation for mathematics today. In fact, just a handful of axioms, called the
Zermelo-Fraenkel with Choice axioms (ZFC), together with a few logical deduction
rules, appear to be sufficient to derive essentially all of mathematics. We’ll examine
these in Chapter 8.
1.4
Our Axioms
The ZFC axioms are important in studying and justifying the foundations of mathematics, but for practical purposes, they are much too primitive. Proving theorems
in ZFC is a little like writing programs in byte code instead of a full-fledged programming language—by one reckoning, a formal proof in ZFC that 2 C 2 D 4
requires more than 20,000 steps! So instead of starting with ZFC, we’re going to
take a huge set of axioms as our foundation: we’ll accept all familiar facts from
high school math.
This will give us a quick launch, but you may find this imprecise specification
of the axioms troubling at times. For example, in the midst of a proof, you may
start to wonder, “Must I prove this little fact or can I take it as an axiom?” There
really is no absolute answer, since what’s reasonable to assume and what requires
proof depends on the circumstances and the audience. A good general guideline is
simply to be up front about what you’re assuming.
“mcs” — 2017/6/5 — 19:42 — page 10 — #18
10
Chapter 1
1.4.1
What is a Proof?
Logical Deductions
Logical deductions, or inference rules, are used to prove new propositions using
previously proved ones.
A fundamental inference rule is modus ponens. This rule says that a proof of P
together with a proof that P IMPLIES Q is a proof of Q.
Inference rules are sometimes written in a funny notation. For example, modus
ponens is written:
Rule.
P;
P IMPLIES Q
Q
When the statements above the line, called the antecedents, are proved, then we
can consider the statement below the line, called the conclusion or consequent, to
also be proved.
A key requirement of an inference rule is that it must be sound: an assignment
of truth values to the letters P , Q, . . . , that makes all the antecedents true must
also make the consequent true. So if we start off with true axioms and apply sound
inference rules, everything we prove will also be true.
There are many other natural, sound inference rules, for example:
Rule.
P IMPLIES Q; Q IMPLIES R
P IMPLIES R
Rule.
NOT .P / IMPLIES NOT .Q/
Q IMPLIES P
On the other hand,
Non-Rule.
NOT .P / IMPLIES NOT .Q/
P IMPLIES Q
is not sound: if P is assigned T and Q is assigned F, then the antecedent is true
and the consequent is not.
As with axioms, we will not be too formal about the set of legal inference rules.
Each step in a proof should be clear and “logical”; in particular, you should state
what previously proved facts are used to derive each new conclusion.
“mcs” — 2017/6/5 — 19:42 — page 11 — #19
1.5. Proving an Implication
1.4.2
11
Patterns of Proof
In principle, a proof can be any sequence of logical deductions from axioms and
previously proved statements that concludes with the proposition in question. This
freedom in constructing a proof can seem overwhelming at first. How do you even
start a proof?
Here’s the good news: many proofs follow one of a handful of standard templates. Each proof has it own details, of course, but these templates at least provide
you with an outline to fill in. We’ll go through several of these standard patterns,
pointing out the basic idea and common pitfalls and giving some examples. Many
of these templates fit together; one may give you a top-level outline while others
help you at the next level of detail. And we’ll show you other, more sophisticated
proof techniques later on.
The recipes below are very specific at times, telling you exactly which words to
write down on your piece of paper. You’re certainly free to say things your own
way instead; we’re just giving you something you could say so that you’re never at
a complete loss.
1.5
Proving an Implication
Propositions of the form “If P , then Q” are called implications. This implication
is often rephrased as “P IMPLIES Q.”
Here are some examples:
(Quadratic Formula) If ax 2 C bx C c D 0 and a ¤ 0, then
p
xD
b ˙ b 2 4ac =2a:
(Goldbach’s Conjecture 1.1.6 rephrased) If n is an even integer greater than
2, then n is a sum of two primes.
If 0 x 2, then x 3 C 4x C 1 > 0.
There are a couple of standard methods for proving an implication.
1.5.1
Method #1
In order to prove that P IMPLIES Q:
1. Write, “Assume P .”
2. Show that Q logically follows.
“mcs” — 2017/6/5 — 19:42 — page 12 — #20
12
Chapter 1
What is a Proof?
Example
Theorem 1.5.1. If 0 x 2, then x 3 C 4x C 1 > 0.
Before we write a proof of this theorem, we have to do some scratchwork to
figure out why it is true.
The inequality certainly holds for x D 0; then the left side is equal to 1 and
1 > 0. As x grows, the 4x term (which is positive) initially seems to have greater
magnitude than x 3 (which is negative). For example, when x D 1, we have
4x D 4, but x 3 D 1 only. In fact, it looks like x 3 doesn’t begin to dominate
until x > 2. So it seems the x 3 C 4x part should be nonnegative for all x between
0 and 2, which would imply that x 3 C 4x C 1 is positive.
So far, so good. But we still have to replace all those “seems like” phrases with
solid, logical arguments. We can get a better handle on the critical x 3 C 4x part
by factoring it, which is not too hard:
x 3 C 4x D x.2
x/.2 C x/
Aha! For x between 0 and 2, all of the terms on the right side are nonnegative. And
a product of nonnegative terms is also nonnegative. Let’s organize this blizzard of
observations into a clean proof.
Proof. Assume 0 x 2. Then x, 2 x and 2Cx are all nonnegative. Therefore,
the product of these terms is also nonnegative. Adding 1 to this product gives a
positive number, so:
x.2 x/.2 C x/ C 1 > 0
Multiplying out on the left side proves that
x 3 C 4x C 1 > 0
as claimed.
There are a couple points here that apply to all proofs:
You’ll often need to do some scratchwork while you’re trying to figure out
the logical steps of a proof. Your scratchwork can be as disorganized as you
like—full of dead-ends, strange diagrams, obscene words, whatever. But
keep your scratchwork separate from your final proof, which should be clear
and concise.
Proofs typically begin with the word “Proof” and end with some sort of delimiter like or “QED.” The only purpose for these conventions is to clarify
where proofs begin and end.
“mcs” — 2017/6/5 — 19:42 — page 13 — #21
1.6. Proving an “If and Only If”
1.5.2
13
Method #2 - Prove the Contrapositive
An implication (“P IMPLIES Q”) is logically equivalent to its contrapositive
NOT .Q/ IMPLIES NOT .P / :
Proving one is as good as proving the other, and proving the contrapositive is sometimes easier than proving the original statement. If so, then you can proceed as
follows:
1. Write, “We prove the contrapositive:” and then state the contrapositive.
2. Proceed as in Method #1.
Example
Theorem 1.5.2. If r is irrational, then
p
r is also irrational.
A number is rational when it equals a quotient of integers —that is, if it equals
m=n for some integers m and n. If it’s not rational, then it’s called irrational. So
p
we must show that if r is not a ratio of integers, then r is also not a ratio of
integers. That’s pretty convoluted! We can eliminate both not’s and simplify the
proof by using the contrapositive instead.
p
Proof. We prove the contrapositive: if r is rational, then r is rational.
p
Assume that r is rational. Then there exist integers m and n such that:
p
m
rD
n
Squaring both sides gives:
m2
rD 2
n
2
2
Since m and n are integers, r is also rational.
1.6
Proving an “If and Only If”
Many mathematical theorems assert that two statements are logically equivalent;
that is, one holds if and only if the other does. Here is an example that has been
known for several thousand years:
Two triangles have the same side lengths if and only if two side lengths
and the angle between those sides are the same.
The phrase “if and only if” comes up so often that it is often abbreviated “iff.”
“mcs” — 2017/6/5 — 19:42 — page 14 — #22
14
Chapter 1
1.6.1
What is a Proof?
Method #1: Prove Each Statement Implies the Other
The statement “P IFF Q” is equivalent to the two statements “P IMPLIES Q” and
“Q IMPLIES P .” So you can prove an “iff” by proving two implications:
1. Write, “We prove P implies Q and vice-versa.”
2. Write, “First, we show P implies Q.” Do this by one of the methods in
Section 1.5.
3. Write, “Now, we show Q implies P .” Again, do this by one of the methods
in Section 1.5.
1.6.2
Method #2: Construct a Chain of Iffs
In order to prove that P is true iff Q is true:
1. Write, “We construct a chain of if-and-only-if implications.”
2. Prove P is equivalent to a second statement which is equivalent to a third
statement and so forth until you reach Q.
This method sometimes requires more ingenuity than the first, but the result can be
a short, elegant proof.
Example
The standard deviation of a sequence of values x1 ; x2 ; : : : ; xn is defined to be:
s
.x1 /2 C .x2 /2 C C .xn /2
(1.3)
n
where is the average or mean of the values:
x1 C x2 C C xn
n
Theorem 1.6.1. The standard deviation of a sequence of values x1 ; : : : ; xn is zero
iff all the values are equal to the mean.
WWD
For example, the standard deviation of test scores is zero if and only if everyone
scored exactly the class average.
Proof. We construct a chain of “iff” implications, starting with the statement that
the standard deviation (1.3) is zero:
s
.x1 /2 C .x2 /2 C C .xn /2
D 0:
(1.4)
n
“mcs” — 2017/6/5 — 19:42 — page 15 — #23
1.7. Proof by Cases
15
Now since zero is the only number whose square root is zero, equation (1.4) holds
iff
.x1 /2 C .x2 /2 C C .xn /2 D 0:
(1.5)
Squares of real numbers are always nonnegative, so every term on the left-hand
side of equation (1.5) is nonnegative. This means that (1.5) holds iff
Every term on the left-hand side of (1.5) is zero.
But a term .xi
(1.6)
/2 is zero iff xi D , so (1.6) is true iff
Every xi equals the mean.
1.7
Proof by Cases
Breaking a complicated proof into cases and proving each case separately is a common, useful proof strategy. Here’s an amusing example.
Let’s agree that given any two people, either they have met or not. If every pair
of people in a group has met, we’ll call the group a club. If every pair of people in
a group has not met, we’ll call it a group of strangers.
Theorem. Every collection of 6 people includes a club of 3 people or a group of 3
strangers.
Proof. The proof is by case analysis5 . Let x denote one of the six people. There
are two cases:
1. Among 5 other people besides x, at least 3 have met x.
2. Among the 5 other people, at least 3 have not met x.
Now, we have to be sure that at least one of these two cases must hold,6 but that’s
easy: we’ve split the 5 people into two groups, those who have shaken hands with
x and those who have not, so one of the groups must have at least half the people.
Case 1: Suppose that at least 3 people did meet x.
This case splits into two subcases:
5 Describing
your approach at the outset helps orient the reader.
of a case analysis argument is showing that you’ve covered all the cases. This is often
obvious, because the two cases are of the form “P ” and “not P .” However, the situation above is not
stated quite so simply.
6 Part
“mcs” — 2017/6/5 — 19:42 — page 16 — #24
16
Chapter 1
What is a Proof?
Case 1.1: No pair among those people met each other. Then these
people are a group of at least 3 strangers. The theorem holds in this
subcase.
Case 1.2: Some pair among those people have met each other. Then
that pair, together with x, form a club of 3 people. So the theorem
holds in this subcase.
This implies that the theorem holds in Case 1.
Case 2: Suppose that at least 3 people did not meet x.
This case also splits into two subcases:
Case 2.1: Every pair among those people met each other. Then these
people are a club of at least 3 people. So the theorem holds in this
subcase.
Case 2.2: Some pair among those people have not met each other.
Then that pair, together with x, form a group of at least 3 strangers. So
the theorem holds in this subcase.
This implies that the theorem also holds in Case 2, and therefore holds in all cases.
1.8
Proof by Contradiction
In a proof by contradiction, or indirect proof, you show that if a proposition were
false, then some false fact would be true. Since a false fact by definition can’t be
true, the proposition must be true.
Proof by contradiction is always a viable approach. However, as the name suggests, indirect proofs can be a little convoluted, so direct proofs are generally preferable when they are available.
Method: In order to prove a proposition P by contradiction:
1. Write, “We use proof by contradiction.”
2. Write, “Suppose P is false.”
3. Deduce something known to be false (a logical contradiction).
4. Write, “This is a contradiction. Therefore, P must be true.”
“mcs” — 2017/6/5 — 19:42 — page 17 — #25
1.9. Good Proofs in Practice
17
Example
p
We’ll prove by contradiction that 2 is irrational. Remember that a number is rational if it is equal to a ratio of integers—for example, 3:5 D 7=2 and 0:1111 D
1=9 are rational numbers.
p
Theorem 1.8.1. 2 is irrational.
p
Proof. We use proof byp
contradiction. Suppose the claim is false, and 2 is rational. Then we can write 2 as a fraction n=d in lowest terms.
Squaring both sides gives 2 D n2 =d 2 and so 2d 2 D n2 . This implies that n is a
multiple of 2 (see Problems 1.15 and 1.16). Therefore n2 must be a multiple of 4.
But since 2d 2 D n2 , we know 2d 2 is a multiple of 4 and so d 2 is a multiple of 2.
This implies that d is a multiple of 2.
So, the numerator and denominator havep
2 as a common factor, which contradicts
the fact that n=d is in lowest terms. Thus, 2 must be irrational.
1.9
Good Proofs in Practice
One purpose of a proof is to establish the truth of an assertion with absolute certainty, and mechanically checkable proofs of enormous length or complexity can
accomplish this. But humanly intelligible proofs are the only ones that help someone understand the subject. Mathematicians generally agree that important mathematical results can’t be fully understood until their proofs are understood. That is
why proofs are an important part of the curriculum.
To be understandable and helpful, more is required of a proof than just logical
correctness: a good proof must also be clear. Correctness and clarity usually go
together; a well-written proof is more likely to be a correct proof, since mistakes
are harder to hide.
In practice, the notion of proof is a moving target. Proofs in a professional
research journal are generally unintelligible to all but a few experts who know all
the terminology and prior results used in the proof. Conversely, proofs in the first
weeks of a beginning course like 6.042 would be regarded as tediously long-winded
by a professional mathematician. In fact, what we accept as a good proof later in
the term will be different from what we consider good proofs in the first couple
of weeks of 6.042. But even so, we can offer some general tips on writing good
proofs:
State your game plan. A good proof begins by explaining the general line of reasoning, for example, “We use case analysis” or “We argue by contradiction.”
“mcs” — 2017/6/5 — 19:42 — page 18 — #26
18
Chapter 1
What is a Proof?
Keep a linear flow. Sometimes proofs are written like mathematical mosaics, with
juicy tidbits of independent reasoning sprinkled throughout. This is not good.
The steps of an argument should follow one another in an intelligible order.
A proof is an essay, not a calculation. Many students initially write proofs the way
they compute integrals. The result is a long sequence of expressions without
explanation, making it very hard to follow. This is bad. A good proof usually
looks like an essay with some equations thrown in. Use complete sentences.
Avoid excessive symbolism. Your reader is probably good at understanding words,
but much less skilled at reading arcane mathematical symbols. Use words
where you reasonably can.
Revise and simplify. Your readers will be grateful.
Introduce notation thoughtfully. Sometimes an argument can be greatly simplified by introducing a variable, devising a special notation, or defining a new
term. But do this sparingly, since you’re requiring the reader to remember
all that new stuff. And remember to actually define the meanings of new
variables, terms, or notations; don’t just start using them!
Structure long proofs. Long programs are usually broken into a hierarchy of smaller
procedures. Long proofs are much the same. When your proof needed facts
that are easily stated, but not readily proved, those fact are best pulled out
as preliminary lemmas. Also, if you are repeating essentially the same argument over and over, try to capture that argument in a general lemma, which
you can cite repeatedly instead.
Be wary of the “obvious.” When familiar or truly obvious facts are needed in a
proof, it’s OK to label them as such and to not prove them. But remember
that what’s obvious to you may not be—and typically is not—obvious to
your reader.
Most especially, don’t use phrases like “clearly” or “obviously” in an attempt
to bully the reader into accepting something you’re having trouble proving.
Also, go on the alert whenever you see one of these phrases in someone else’s
proof.
Finish. At some point in a proof, you’ll have established all the essential facts
you need. Resist the temptation to quit and leave the reader to draw the
“obvious” conclusion. Instead, tie everything together yourself and explain
why the original claim follows.
“mcs” — 2017/6/5 — 19:42 — page 19 — #27
1.10. References
19
Creating a good proof is a lot like creating a beautiful work of art. In fact,
mathematicians often refer to really good proofs as being “elegant” or “beautiful.”
It takes a practice and experience to write proofs that merit such praises, but to
get you started in the right direction, we will provide templates for the most useful
proof techniques.
Throughout the text there are also examples of bogus proofs—arguments that
look like proofs but aren’t. Sometimes a bogus proof can reach false conclusions
because of missteps or mistaken assumptions. More subtle bogus proofs reach
correct conclusions, but do so in improper ways such as circular reasoning, leaping
to unjustified conclusions, or saying that the hard part of the proof is “left to the
reader.” Learning to spot the flaws in improper proofs will hone your skills at seeing
how each proof step follows logically from prior steps. It will also enable you to
spot flaws in your own proofs.
The analogy between good proofs and good programs extends beyond structure.
The same rigorous thinking needed for proofs is essential in the design of critical computer systems. When algorithms and protocols only “mostly work” due
to reliance on hand-waving arguments, the results can range from problematic to
catastrophic. An early example was the Therac 25, a machine that provided radiation therapy to cancer victims, but occasionally killed them with massive overdoses
due to a software race condition. A further example of a dozen years ago (August
2004) involved a single faulty command to a computer system used by United and
American Airlines that grounded the entire fleet of both companies—and all their
passengers!
It is a certainty that we’ll all one day be at the mercy of critical computer systems
designed by you and your classmates. So we really hope that you’ll develop the
ability to formulate rock-solid logical arguments that a system actually does what
you think it should do!
1.10
References
[14], [1], [49], [18], [22]
“mcs” — 2017/6/5 — 19:42 — page 20 — #28
20
Chapter 1
What is a Proof?
Problems for Section 1.1
Class Problems
Problem 1.1.
Albert announces to his class that he plans to surprise them with a quiz sometime
next week.
His students first wonder if the quiz could be on Friday of next week. They
reason that it can’t: if Albert didn’t give the quiz before Friday, then by midnight
Thursday, they would know the quiz had to be on Friday, and so the quiz wouldn’t
be a surprise any more.
Next the students wonder whether Albert could give the surprise quiz Thursday.
They observe that if the quiz wasn’t given before Thursday, it would have to be
given on the Thursday, since they already know it can’t be given on Friday. But
having figured that out, it wouldn’t be a surprise if the quiz was on Thursday either.
Similarly, the students reason that the quiz can’t be on Wednesday, Tuesday, or
Monday. Namely, it’s impossible for Albert to give a surprise quiz next week. All
the students now relax, having concluded that Albert must have been bluffing. And
since no one expects the quiz, that’s why, when Albert gives it on Tuesday next
week, it really is a surprise!
What, if anything, do you think is wrong with the students’ reasoning?
Problem 1.2.
The Pythagorean Theorem says that if a and b are the lengths of the sides of a right
triangle, and c is the length of its hypotenuse, then
a2 C b 2 D c 2 :
This theorem is so fundamental and familiar that we generally take it for granted.
But just being familiar doesn’t justify calling it “obvious”—witness the fact that
people have felt the need to devise different proofs of it for milllenia.7 In this
problem we’ll examine a particularly simple “proof without words” of the theorem.
Here’s the strategy. Suppose you are given four different colored copies of a
right triangle with sides of lengths a, b and c, along with a suitably sized square,
as shown in Figure 1.1.
(a) You will first arrange the square and four triangles so they form a c c square.
From this arrangement you will see that the square is .b a/ .b a/.
7 Over
a hundred different proofs are listed on the mathematics website http://www.cut-theknot.org/pythagoras/.
“mcs” — 2017/6/5 — 19:42 — page 21 — #29
1.10. References
21
c
b
a
Figure 1.1 Right triangles and square.
(b) You will then arrange the same shapes so they form two squares, one a a
and the other b b.
You know that the area of an s s square is s 2 . So appealing to the principle that
Area is Preserved by Rearranging,
you can now conclude that a2 C b 2 D c 2 , as claimed.
This really is an elegant and convincing proof of the Pythagorean Theorem, but it
has some worrisome features. One concern is that there might be something special
about the shape of these particular triangles and square that makes the rearranging
possible—for example, suppose a D b?
(c) How would you respond to this concern?
(d) Another concern is that a number of facts about right triangles, squares and
lines are being implicitly assumed in justifying the rearrangements into squares.
Enumerate some of these assumed facts.
Problem 1.3.
What’s going on here?!
1D
p 2
p
p
p p
1 D . 1/. 1/ D
1
1D
1 D
1:
(a) Precisely identify and explain the mistake(s) in this bogus proof.
(b) Prove (correctly) that if 1 D
1, then 2 D 1.
“mcs” — 2017/6/5 — 19:42 — page 22 — #30
22
Chapter 1
What is a Proof?
(c) Every positive real number r has two square roots, one positive and the other
p
negative. The standard convention is that the expression r refers to the positive
square root of r. Assuming familiar properties of multiplication of real numbers,
prove that for positive real numbers r and s,
p p
p
rs D r s:
Problem 1.4.
Identify exactly where the bugs are in each of the following bogus proofs.8
(a) Bogus Claim: 1=8 > 1=4:
Bogus proof.
3>2
3 log10 .1=2/ > 2 log10 .1=2/
log10 .1=2/3 > log10 .1=2/2
.1=2/3 > .1=2/2 ;
and the claim now follows by the rules for multiplying fractions.
(b) Bogus proof : 1¢ D $0:01 D .$0:1/2 D .10¢/2 D 100¢ D $1: (c) Bogus Claim: If a and b are two equal real numbers, then a D 0.
Bogus proof.
aDb
a2 D ab
a2
.a
b 2 D ab
b2
b/.a C b/ D .a
b/b
aCb Db
a D 0:
8 From
[48], Twenty Years Before the Blackboard by Michael Stueben and Diane Sandford
“mcs” — 2017/6/5 — 19:42 — page 23 — #31
1.10. References
23
Problem 1.5.
It’s a fact that the Arithmetic Mean is at least as large as the Geometric Mean,
namely,
aCb p
ab
2
for all nonnegative real numbers a and b. But there’s something objectionable
about the following proof of this fact. What’s the objection, and how would you fix
it?
Bogus proof.
aCb ‹ p
ab;
2
‹ p
a C b 2 ab;
‹
a2 C 2ab C b 2 4ab;
a2
‹
2ab C b 2 0;
.a
2
b/ 0
so
so
so
so
which we know is true.
The last statement is true because a b is a real number, and the square of a real
number is never negative. This proves the claim.
Practice Problems
Problem 1.6.
Why does the “surprise” paradox of Problem 1.1 present a philosophical problem
but not a mathematical one?
Problems for Section 1.5
Homework Problems
Problem 1.7.
Show that log7 n is either an integer or irrational, where n is a positive integer. Use
whatever familiar facts about integers and primes you need, but explicitly state such
facts.
“mcs” — 2017/6/5 — 19:42 — page 24 — #32
24
Chapter 1
What is a Proof?
Problems for Section 1.7
Practice Problems
Problem 1.8.
Prove by cases that
max.r; s/ C min.r; s/ D r C s
(*)
for all real numbers r; s.
Class Problems
Problem 1.9.
If we raise an irrational number to p
an irrational power, can the result be rational?
p 2
Show that it can by considering 2 and arguing by cases.
Problem 1.10.
Prove by cases that
jr C sj jrj C jsj
(1)
for all real numbers r; s.9
Homework Problems
Problem 1.11. (a) Suppose that
a C b C c D d;
where a; b; c; d are nonnegative integers.
Let P be the assertion that d is even. Let W be the assertion that exactly one among
a; b; c are even, and let T be the assertion that all three are even.
Prove by cases that
P IFF ŒW OR T :
(b) Now suppose that
w2 C x2 C y 2 D z2;
9 The
absolute value jrj of r equals whichever of r or r is not negative.
“mcs” — 2017/6/5 — 19:42 — page 25 — #33
1.10. References
25
where w; x; y; z are nonnegative integers. Let P be the assertion that z is even, and
let R be the assertion that all three of w; x; y are even. Prove by cases that
P IFF R:
Hint: An odd number equals 2m C 1 for some integer m, so its square equals
4.m2 C m/ C 1.
Exam Problems
Problem 1.12.
p
Prove that there is an irrational
number a such that a 3 is rational.
p
p 3
Hint: Consider 3 2 and argue by cases.
Problems for Section 1.8
Practice Problems
Problem 1.13.
Prove that for any n > 0, if an is even, then a is even.
Hint: Contradiction.
Problem 1.14.
p
Prove that if a b D n, then either a or b must be n, where a; b, and n are
nonnegative real numbers. Hint: by contradiction, Section 1.8.
Problem 1.15.
Let n be a nonnegative integer.
(a) Explain why if n2 is even—that is, a multiple of 2—then n is even.
(b) Explain why if n2 is a multiple of 3, then n must be a multiple of 3.
Problem 1.16.
Give an example of two distinct positive integers m; n such that n2 is a multiple of
m, but n is not a multiple of m. How about having m be less than n?
“mcs” — 2017/6/5 — 19:42 — page 26 — #34
26
Chapter 1
What is a Proof?
Class Problems
Problem 1.17.
p
How far can you generalize
the
proof
of
Theorem
1.8.1
that
2 is irrational? For
p
example, how about 3?
Problem 1.18.
Prove that log4 6 is irrational.
Problem 1.19.
p
p
Prove by p
contradiction
that
3
C
2 is irrational.
p p
p
Hint: . 3 C 2/. 3
2/
Problem 1.20.
Here is a generalization of Problem 1.17 that you may not have thought of:
Lemma. Let the coefficients of the polynomial
a0 C a1 x C a2 x 2 C C am
1x
m 1
C xm
be integers. Then any real root of the polynomial is either integral or irrational.
p
m
(a) Explain why the Lemma immediately implies that k is irrational whenever
k is not an mth power of some integer.
(b) Carefully prove the Lemma.
You may find it helpful to appeal to:
Fact. If a prime p is a factor of some power of an integer, then it is a factor of that
integer.
You may assume this Fact without writing down its proof, but see if you can explain
why it is true.
Exam Problems
Problem 1.21.
Prove that log9 12 is irrational.
“mcs” — 2017/6/5 — 19:42 — page 27 — #35
1.10. References
27
Problem 1.22.
Prove that log12 18 is irrational.
Problem 1.23.
p
3
A familiar proof that 72 is irrational depends on the fact that a certain equation
among those below is unsatisfiable by integers a; b > 0. Note that more than one
is unsatisfiable. Indicate the equation that would
appear in the proof, and explain
p
3
2
why it is unsatisfiable. (Do not assume that 7 is irrational.)
i. a2 D 72 C b 2
ii. a3 D 72 C b 3
iii. a2 D 72 b 2
iv. a3 D 72 b 3
v. a3 D 73 b 3
vi. .ab/3 D 72
Homework Problems
Problem 1.24.
The fact that that there are irrational numbers a; b such that ab is rational was
proved in Problem 1.9 by cases. Unfortunately, that proof was nonconstructive: it
didn’t reveal
p a specific pair a; b with this property. But in fact, it’s easy to do this:
let a WWD 2 and bpWWD 2 log2 3.
We know a D 2 is irrational, and ab D 3 by definition. Finish the proof that
these values for a; b work, by showing that 2 log2 3 is irrational.
Problem 1.25.
p
Here is a different proof that 2 is irrational, taken from the American Mathematical Monthly, v.116, #1, Jan. 2009, p.69:
p
Proof. Suppose for the sake ofcontradiction
that 2 is rational, and choose the
p
least integer q > 0 such that
2 1 q is a nonnegative integer. Let q 0 WWD
p
p
2 1 q. Clearly 0 < q 0 < q. But an easy computation shows that
2 1 q0
is a nonnegative integer, contradicting the minimality of q.
“mcs” — 2017/6/5 — 19:42 — page 28 — #36
28
Chapter 1
What is a Proof?
(a) This proof was written for an audience of college teachers, and at this point it
is a little more concise than desirable. Write out a more complete version which
includes an explanation of each step.
(b) Now that you have justified the steps in this proof, do you have a preference
for one of these proofs over the other? Why? Discuss these questions with your
teammates for a few minutes and summarize your team’s answers on your whiteboard.
Problem 1.26.
For n D 40, the value of polynomial p.n/ WWD n2 C n C 41 is not prime, as noted
in Section 1.1. But we could have predicted based on general principles that no
nonconstant polynomial can generate only prime numbers.
In particular, let q.n/ be a polynomial with integer coefficients, and let c WWD q.0/
be the constant term of q.
(a) Verify that q.cm/ is a multiple of c for all m 2 Z.
(b) Show that if q is nonconstant and c > 1, then as n ranges over the nonnegative
integers N there are infinitely many q.n/ 2 Z that are not primes.
Hint: You may assume the familiar fact that the magnitude of any nonconstant
polynomial q.n/ grows unboundedly as n grows.
(c) Conclude that for every nonconstant polynomial q there must be an n 2 N
such that q.n/ is not prime. Hint: Only one easy case remains.
“mcs” — 2017/6/5 — 19:42 — page 29 — #37
2
The Well Ordering Principle
Every nonempty set of nonnegative integers has a smallest element.
This statement is known as The Well Ordering Principle (WOP). Do you believe
it? Seems sort of obvious, right? But notice how tight it is: it requires a nonempty
set—it’s false for the empty set which has no smallest element because it has no
elements at all. And it requires a set of nonnegative integers—it’s false for the
set of negative integers and also false for some sets of nonnegative rationals—for
example, the set of positive rationals. So, the Well Ordering Principle captures
something special about the nonnegative integers.
While the Well Ordering Principle may seem obvious, it’s hard to see offhand
why it is useful. But in fact, it provides one of the most important proof rules in
discrete mathematics. In this chapter, we’ll illustrate the power of this proof method
with a few simple examples.
2.1
Well Ordering Proofs
We actually
have already taken the Well Ordering Principle for granted in proving
p
that 2 is irrational. That proof assumed that for any positive integers m and n,
the fraction m=n can be written in lowest terms, that is, in the form m0 =n0 where
m0 and n0 are positive integers with no common prime factors. How do we know
this is always possible?
Suppose to the contrary that there are positive integers m and n such that the
fraction m=n cannot be written in lowest terms. Now let C be the set of positive
integers that are numerators of such fractions. Then m 2 C , so C is nonempty. By
WOP, there must be a smallest integer m0 2 C . So by definition of C , there is an
integer n0 > 0 such that
the fraction
m0
cannot be written in lowest terms.
n0
This means that m0 and n0 must have a common prime factor, p > 1. But
m0 =p
m0
D
;
n0 =p
n0
“mcs” — 2017/6/5 — 19:42 — page 30 — #38
30
Chapter 2
The Well Ordering Principle
so any way of expressing the left-hand fraction in lowest terms would also work for
m0 =n0 , which implies
the fraction
m0 =p
cannot be in written in lowest terms either.
n0 =p
So by definition of C , the numerator m0 =p is in C . But m0 =p < m0 , which
contradicts the fact that m0 is the smallest element of C .
Since the assumption that C is nonempty leads to a contradiction, it follows that
C must be empty. That is, that there are no numerators of fractions that can’t be
written in lowest terms, and hence there are no such fractions at all.
We’ve been using the Well Ordering Principle on the sly from early on!
2.2
Template for WOP Proofs
More generally, there is a standard way to use Well Ordering to prove that some
property, P .n/ holds for every nonnegative integer n. Here is a standard way to
organize such a well ordering proof:
To prove that “P .n/ is true for all n 2 N” using the Well Ordering Principle:
Define the set C of counterexamples to P being true. Specifically, define
C WWD fn 2 N j NOT.P .n// is trueg:
(The notation fn j Q.n/g means “the set of all elements n for which Q.n/
is true.” See Section 4.1.4.)
Assume for proof by contradiction that C is nonempty.
By WOP, there will be a smallest element n in C .
Reach a contradiction somehow—often by showing that P .n/ is actually
true or by showing that there is another member of C that is smaller than
n. This is the open-ended part of the proof task.
Conclude that C must be empty, that is, no counterexamples exist.
2.2.1
Summing the Integers
Let’s use this template to prove
“mcs” — 2017/6/5 — 19:42 — page 31 — #39
2.2. Template for WOP Proofs
31
Theorem 2.2.1.
1 C 2 C 3 C C n D n.n C 1/=2
(2.1)
for all nonnegative integers n.
First, we’d better address a couple of ambiguous special cases before they trip us
up:
If n D 1, then there is only one term in the summation, and so 1 C 2 C 3 C
C n is just the term 1. Don’t be misled by the appearance of 2 and 3 or by
the suggestion that 1 and n are distinct terms!
If n D 0, then there are no terms at all in the summation. By convention, the
sum in this case is 0.
So, while the three dots notation, which is called an ellipsis, is convenient, you
have to watch out for these special cases where the notation is misleading. In
fact, whenever you see an ellipsis, you should be on the lookout to be sure you
understand the pattern, watching out for the beginning and the end.
We could have eliminated the need for guessing by rewriting the left side of (2.1)
with summation notation:
n
X
i D1
i
or
X
i:
1i n
Both of these expressions denote the sum of all values taken by the expression to
the right of the sigma as the variable i ranges from 1 to n. Both expressions make
it clear what (2.1) means when n D 1. The second expression makes it clear that
when n D 0, there are no terms in the sum, though you still have to know the
convention that a sum of no numbers equals 0 (the product of no numbers is 1, by
the way).
OK, back to the proof:
Proof. By contradiction. Assume that Theorem 2.2.1 is false. Then, some nonnegative integers serve as counterexamples to it. Let’s collect them in a set:
C WWD fn 2 N j 1 C 2 C 3 C C n ¤
n.n C 1/
g:
2
Assuming there are counterexamples, C is a nonempty set of nonnegative integers.
So, by WOP, C has a minimum element, which we’ll call c. That is, among the
nonnegative integers, c is the smallest counterexample to equation (2.1).
Since c is the smallest counterexample, we know that (2.1) is false for n D c but
true for all nonnegative integers n < c. But (2.1) is true for n D 0, so c > 0. This
“mcs” — 2017/6/5 — 19:42 — page 32 — #40
32
Chapter 2
The Well Ordering Principle
means c 1 is a nonnegative integer, and since it is less than c, equation (2.1) is
true for c 1. That is,
1 C 2 C 3 C C .c
1/ D
.c
1/c
2
:
But then, adding c to both sides, we get
1 C 2 C 3 C C .c
1/ C c D
.c
1/c
2
Cc D
c2
c C 2c
c.c C 1/
D
;
2
2
which means that (2.1) does hold for c, after all! This is a contradiction, and we
are done.
2.3
Factoring into Primes
We’ve previously taken for granted the Prime Factorization Theorem, also known
as the Unique Factorization Theorem and the Fundamental Theorem of Arithmetic,
which states that every integer greater than one has a unique1 expression as a product of prime numbers. This is another of those familiar mathematical facts which
are taken for granted but are not really obvious on closer inspection. We’ll prove
the uniqueness of prime factorization in a later chapter, but well ordering gives an
easy proof that every integer greater than one can be expressed as some product of
primes.
Theorem 2.3.1. Every positive integer greater than one can be factored as a product of primes.
Proof. The proof is by WOP.
Let C be the set of all integers greater than one that cannot be factored as a
product of primes. We assume C is not empty and derive a contradiction.
If C is not empty, there is a least element n 2 C by WOP. This n can’t be prime,
because a prime by itself is considered a (length one) product of primes, and no
such products are in C .
So n must be a product of two integers a and b where 1 < a; b < n. Since a
and b are smaller than the smallest element in C , we know that a; b … C . In other
words, a can be written as a product of primes p1 p2 pk and b as a product of
primes q1 ql . Therefore, n D p1 pk q1 ql can be written as a product of
primes, contradicting the claim that n 2 C . Our assumption that C is not empty
must therefore be false.
1 . . . unique
up to the order in which the prime factors appear
“mcs” — 2017/6/5 — 19:42 — page 33 — #41
2.4. Well Ordered Sets
2.4
33
Well Ordered Sets
A set of real numbers is well ordered when each of its nonempty subsets has a
minimum element. The Well Ordering Principle says that the set of nonnegative
integers is well ordered, but so are lots of other sets of real numbers according to
this more general form of WOP. A simple example would be the set rN of numbers
of the form rn, where r is a positive real number and n 2 N. (Why does this work?)
Well ordering commonly comes up in computer science as a method for proving
that computations won’t run forever. The idea is to assign a value to each successive
step of a computation so that the values get smaller at every step. If the values are
all from a well ordered set, then the computation can’t run forever, because if it did,
the values assigned to its successive steps would define a subset with no minimum
element. You’ll see several examples of this technique applied in Chapter 6 to prove
that various state machines will eventually terminate.
Notice that a set may have a minimum element but not be well ordered. The set
of nonnegative rational numbers is an example: it has a minimum element zero,
but it also has nonempty subsets that don’t have minimum elements—the positive
rationals, for example.
The following theorem is a tiny generalization of the Well Ordering Principle.
Theorem 2.4.1. For any nonnegative integer n the set of integers greater than or
equal to n is well ordered.
This theorem is just as obvious as the Well Ordering Principle, and it would
be harmless to accept it as another axiom. But repeatedly introducing axioms gets
worrisome after a while, and it’s worth noticing when a potential axiom can actually
be proved. This time we can easily prove Theorem 2.4.1 using the Well Ordering
Principle:
Proof. Let S be any nonempty set of integers n. Now add n to each of the
elements in S ; let’s call this new set S C n. Now S C n is a nonempty set of
nonnegative integers, and so by the Well Ordering Principle, it has a minimum
element m. But then it’s easy to see that m n is the minimum element of S . The definition of well ordering states that every subset of a well ordered set
is well ordered, and this yields two convenient, immediate corollaries of Theorem 2.4.1:
Definition 2.4.2. A lower bound (respectively, upper bound) for a set S of real
numbers is a number b such that b s (respectively, b s) for every s 2 S .
“mcs” — 2017/6/5 — 19:42 — page 34 — #42
34
Chapter 2
The Well Ordering Principle
Note that a lower or upper bound of set S is not required to be in the set.
Corollary 2.4.3. Any set of integers with a lower bound is well ordered.
Proof. A set of integers with a lower bound b 2 R will also have the integer n D
bbc as a lower bound, where bbc, called the floor of b, is gotten by rounding down
b to the nearest integer. So Theorem 2.4.1 implies the set is well ordered.
Corollary 2.4.3 leads to another principle we usually tahe for granted:
Corollary 2.4.4. Any nonempty set of integers with an upper bound has a maximum
element.
Proof. Suppose a set S of integers has an upper bound b 2 R. Now multiply each
element of S by -1; let’s call this new set of elements S . Now, of course, b is a
lower bound of S . So S has a minimum element m by Corollary 2.4.3. But
then it’s easy to see that m is the maximum element of S.
Finite sets are yet another routine example of well ordered set.
Lemma 2.4.5. Every nonempty finite set of real numbers is well ordered.
Proof. Since subsets of finite sets are finite, it is sufficient to prove that every finite
set has a minimum element.
We prove this using the WOP on the size of finite sets.
Let C be the set of positive integers n such that some set of size n has no minimum element. Assume for the sake of contradiction that C is nonempty. By WOP,
there is a minimum integer m 2 C .
Every set of size one obviously has a minimum element, so m 2.
Now let F be a set of m real numbers. We will reach a contradiction by showing
that F has a minimum element.
So let r0 be an element of F . Since m 2, removing r0 from F leaves a
nonempty set F 0 smaller than m. Since m is the smallest element of C , we know
F 0 has a minimum element r1 . But that means the smaller of r0 and r1 is the
minimum element of F .
2.4.1
A Different Well Ordered Set (Optional)
The set F of fractions that can be expressed in the form n=.n C 1/:
0 1 2 3
n
; ; ; ;:::;
;:::;
1 2 3 4
nC1
is well ordered. The minimum element of any nonempty subset of F is simply the
one with the minimum numerator when expressed in the form n=.n C 1/.
“mcs” — 2017/6/5 — 19:42 — page 35 — #43
2.4. Well Ordered Sets
35
Now we can define a very different well ordered set by adding nonnegative integers to numbers in F. That is, we take all the numbers of the form n C f where n is
a nonnegative integer and f is a number in F. Let’s call this set of numbers—you
guessed it—N C F. There is a simple recipe for finding the minimum number in
any nonempty subset of N C F, which explains why this set is well ordered:
Lemma 2.4.6. N C F is well ordered.
Proof. Given any nonempty subset S of N C F, look at all the nonnegative integers
n such that n C f is in S for some f 2 F. This is a nonempty set nonnegative
integers, so by the WOP, there is a minimum such integer; call it ns .
By definition of ns , there is some f 2 F such that nS C f is in the set S. So
the set all fractions f such that nS C f 2 S is a nonempty subset of F, and since
F is well ordered, this nonempty set contains a minimum element; call it fS . Now
it easy to verify that nS C fS is the minimum element of S (Problem 2.20).
The set N C F is significantly different from the examples above, and it provides a hint of the rich collection of well ordered sets. In all the earlier examples,
each element was greater than only a finite number of other elements. In N C F,
every element greater than or equal to 1 can be the first element in strictly decreasing sequences of elements of arbitrary finite length. For example, the following
decreasing sequences of elements in N C F all start with 1:
1; 0:
1; 12 ; 0:
1; 23 ; 12 ; 0:
1; 34 ; 23 ; 12 ; 0:
::
:
Nevertheless, since N C F is well ordered, it is impossible to find an infinite decreasing sequence of elements in N C F, because the set of elements in such a
sequence would have no minimum.
Problems for Section 2.2
Practice Problems
Problem 2.1.
For practice using the Well Ordering Principle, fill in the template of an easy to
“mcs” — 2017/6/5 — 19:42 — page 36 — #44
36
Chapter 2
The Well Ordering Principle
prove fact: every amount of postage that can be assembled using only 10 cent and
15 cent stamps is divisible by 5.
In particular, let the notation “j j k” indicate that integer j is a divisor of integer
k, and let S.n/ mean that exactly n cents postage can be assembled using only 10
and 15 cent stamps. Then the proof shows that
S.n/ IMPLIES 5 j n;
for all nonnegative integers n:
(2.2)
Fill in the missing portions (indicated by “. . . ”) of the following proof of (2.2).
Let C be the set of counterexamples to (2.2), namely
C WWD fn j : : :g
Assume for the purpose of obtaining a contradiction that C is nonempty.
Then by the WOP, there is a smallest number m 2 C . This m must be
positive because . . . .
But if S.m/ holds and m is positive, then S.m
must hold, because . . . .
So suppose S.m
10/ holds. Then 5 j .m
10/ or S.m
15/
10/, because. . .
But if 5 j .m 10/, then obviously 5 j m, contradicting the fact that m
is a counterexample.
Next, if S.m
15/ holds, we arrive at a contradiction in the same way.
Since we get a contradiction in both cases, we conclude that. . .
which proves that (2.2) holds.
Problem 2.2.
The Fibonacci numbers F .0/; F .1/; F .2/; : : : are defined as follows:
8
ˆ
if n D 0;
<0
F .n/ WWD 1
if n D 1;
ˆ
:
F .n 1/ C F .n 2/ if n > 1:
Exactly which sentence(s) in the following bogus proof contain logical errors?
Explain.
False Claim. Every Fibonacci number is even.
“mcs” — 2017/6/5 — 19:42 — page 37 — #45
2.4. Well Ordered Sets
37
Bogus proof. Let all the variables n; m; k mentioned below be nonnegative integer
valued.
1. The proof is by the WOP.
2. Let EF.n/ mean that F .n/ is even.
3. Let C be the set of counterexamples to the assertion that EF.n/ holds for all
n 2 N, namely,
C WWD fn 2 N j NOT.EF.n//g:
4. We prove by contradiction that C is empty. So assume that C is not empty.
5. By WOP, there is a least nonnegative integer m 2 C .
6. Then m > 0, since F .0/ D 0 is an even number.
7. Since m is the minimum counterexample, F .k/ is even for all k < m.
8. In particular, F .m
1/ and F .m
2/ are both even.
9. But by the definition, F .m/ equals the sum F .m
even numbers, and so it is also even.
1/ C F .m
2/ of two
10. That is, EF.m/ is true.
11. This contradicts the condition in the definition of m that NOT.EF.m// holds.
12. This contradition implies that C must be empty. Hence, F .n/ is even for all
n 2 N.
Problem 2.3.
In Chapter 2, the Well Ordering Principle was used to show that all positive rational
numbers can be written in “lowest terms,” that is, as a ratio of positive integers with
no common factor prime factor. Below is a different proof which also arrives at this
correct conclusion, but this proof is bogus. Identify every step at which the proof
makes an unjustified inference.
Bogus proof. Suppose to the contrary that there was positive rational q such that q
cannot be written in lowest terms. Now let C be the set of such rational numbers
that cannot be written in lowest terms. Then q 2 C , so C is nonempty. So there
“mcs” — 2017/6/5 — 19:42 — page 38 — #46
38
Chapter 2
The Well Ordering Principle
must be a smallest rational q0 2 C . So since q0 =2 < q0 , it must be possible to
express q0 =2 in lowest terms, namely,
q0
m
D
2
n
(2.3)
for positive integers m; n with no common prime factor. Now we consider two
cases:
Case 1: [n is odd]. Then 2m and n also have no common prime factor, and
therefore
m
2m
q0 D 2 D
n
n
expresses q0 in lowest terms, a contradiction.
Case 2: [n is even]. Any common prime factor of m and n=2 would also be a
common prime factor of m and n. Therefore m and n=2 have no common prime
factor, and so
m
q0 D
n=2
expresses q0 in lowest terms, a contradiction.
Since the assumption that C is nonempty leads to a contradiction, it follows that
C is empty—that is, there are no counterexamples.
Class Problems
Problem 2.4.
Use the Well Ordering Principle 2 to prove that
n
X
kD0
k2 D
n.n C 1/.2n C 1/
:
6
(2.4)
for all nonnegative integers n.
Problem 2.5.
Use the Well Ordering Principle to prove that there is no solution over the positive
integers to the equation:
4a3 C 2b 3 D c 3 :
2 Proofs
by other methods such as induction or by appeal to known formulas for similar sums will
not receive full credit.
“mcs” — 2017/6/5 — 19:42 — page 39 — #47
2.4. Well Ordered Sets
39
Problem 2.6.
You are given a series of envelopes, respectively containing 1; 2; 4; : : : ; 2m dollars.
Define
Property m: For any nonnegative integer less than 2mC1 , there is a
selection of envelopes whose contents add up to exactly that number
of dollars.
Use the Well Ordering Principle (WOP) to prove that Property m holds for all
nonnegative integers m.
Hint: Consider two cases: first, when the target number of dollars is less than
2m and second, when the target is at least 2m .
Homework Problems
Problem 2.7.
Use the Well Ordering Principle to prove that any integer greater than or equal to 8
can be represented as the sum of nonnegative integer multiples of 3 and 5.
Problem 2.8.
Use the Well Ordering Principle to prove that any integer greater than or equal to
50 can be represented as the sum of nonnegative integer multiples of 7, 11, and 13.
Problem 2.9.
Euler’s Conjecture in 1769 was that there are no positive integer solutions to the
equation
a4 C b 4 C c 4 D d 4 :
Integer values for a; b; c; d that do satisfy this equation were first discovered in
1986. So Euler guessed wrong, but it took more than two centuries to demonstrate
his mistake.
Now let’s consider Lehman’s equation, similar to Euler’s but with some coefficients:
8a4 C 4b 4 C 2c 4 D d 4
(2.5)
Prove that Lehman’s equation (2.5) really does not have any positive integer
solutions.
Hint: Consider the minimum value of a among all possible solutions to (2.5).
“mcs” — 2017/6/5 — 19:42 — page 40 — #48
40
Chapter 2
The Well Ordering Principle
Problem 2.10.
Use the Well Ordering Principle to prove that
n 3n=3
(2.6)
for every nonnegative integer n.
Hint: Verify (2.6) for n 4 by explicit calculation.
Problem 2.11.
A winning configuration in the game of Mini-Tetris is a complete tiling of a 2 n
board using only the three shapes shown below:
For example, here are several possible winning configurations on a 2 5 board:
(a) Let Tn denote the number of different winning configurations on a 2n board.
Determine the values of T1 , T2 and T3 .
(b) Express Tn in terms of Tn
1
and Tn
2
for n > 2.
(c) Use the Well Ordering Principle to prove that the number of winning configurations on a 2 n Mini-Tetris board is:3
Tn D
2nC1 C . 1/n
3
(*)
Problem 2.12.
Mini-Tetris is a game whose objective is to provide a complete “tiling” of a 2 n
board using tiles of specified shapes. In this problem we consider the following set
of five tiles:
3A
good question is how someone came up with equation (*) in the first place. A simple Well
Ordering proof gives no hint about this, but it should be absolutely convincing anyway.
“mcs” — 2017/6/5 — 19:42 — page 41 — #49
2.4. Well Ordered Sets
41
For example, there are two possible tilings of a 2 1 board:
Also, here are three tilings for a 2 2 board:
Note that tiles may not be rotated, which is why the second and third of the above
tilings count as different, even though one is a 180o rotation of the other. (A 90o
degree rotation of these shapes would not count as a tiling at all.)
(a) There are four more 2 2 tilings in addition to the three above. What are they?
Let Tn denote the number of different tilings of a 2 n board. We know that
T1 D 2 and T2 D 7. Also, T0 D 1 because there is exactly one way to tile a 2 0
board—don’t use any tiles.
(b) Tn can be specified in terms of Tn
Tn D 2Tn
1
and Tn
1
C 3Tn
for n 2.
Briefly explain how to justify this equation.
2
as follows:
2
(2.7)
“mcs” — 2017/6/5 — 19:42 — page 42 — #50
42
Chapter 2
The Well Ordering Principle
(c) Use the Well Ordering Principle to prove that for n 0, the number Tn of
tilings of a 2 n Mini-Tetris board is:
3nC1 C . 1/n
:
4
Exam Problems
Problem 2.13.
Except for an easily repaired omission, the following proof using the Well Ordering
Principle shows that every amount of postage that can be paid exactly using only
10 cent and 15 cent stamps, is divisible by 5.
Namely, let the notation “j j k” indicate that integer j is a divisor of integer k,
and let S.n/ mean that exactly n cents postage can be assembled using only 10 and
15 cent stamps. Then the proof shows that
S.n/ IMPLIES 5 j n;
for all nonnegative integers n:
(2.8)
Fill in the missing portions (indicated by “. . . ”) of the following proof of (2.8), and
at the end, identify the minor mistake in the proof and how to fix it.
Let C be the set of counterexamples to (2.8), namely
C WWD fn j S.n/ and NOT.5 j n/g
Assume for the purpose of obtaining a contradiction that C is nonempty.
Then by the WOP, there is a smallest number m 2 C . Then S.m 10/
or S.m 15/ must hold, because the m cents postage is made from 10
and 15 cent stamps, so we remove one.
So suppose S.m
But if 5 j .m
10/ holds. Then 5 j .m
10/, because. . .
10/, then 5 j m, because. . .
contradicting the fact that m is a counterexample.
Next suppose S.m 15/ holds. Then the proof for m 10 carries
over directly for m 15 to yield a contradiction in this case as well.
Since we get a contradiction in both cases, we conclude that C must
be empty. That is, there are no counterexamples to (2.8), which proves
that (2.8) holds.
The proof makes an implicit assumption about the value of m. State the assumption and justify it in one sentence.
“mcs” — 2017/6/5 — 19:42 — page 43 — #51
2.4. Well Ordered Sets
43
Problem 2.14. (a) Prove using the Well Ordering Principle that, using 6¢, 14¢, and
21¢ stamps, it is possible to make any amount of postage over 50¢. To save time,
you may specify assume without proof that 50¢, 51¢, . . . 100¢ are all makeable, but
you should clearly indicate which of these assumptions your proof depends on.
(b) Show that 49¢ is not makeable.
Problem 2.15.
We’ll use the Well Ordering Principle to prove that for every positive integer n, the
sum of the first n odd numbers is n2 , that is,
n
X1
.2i C 1/ D n2 ;
(2.9)
i D0
for all n > 0.
Assume to the contrary that equation (2.9) failed for some positive integer n. Let
m be the least such number.
(a) Why must there be such an m?
(b) Explain why m 2.
(c) Explain why part (b) implies that
m
X1
.2.i
1/ C 1/ D .m
1/2 :
(2.10)
i D1
(d) What term should be added to the left-hand side of (2.10) so the result equals
m
X
.2.i
1/ C 1/‹
i D1
(e) Conclude that equation (2.9) holds for all positive integers n.
Problem 2.16.
Use the Well Ordering Principle (WOP) to prove that
2 C 4 C C 2n D n.n C 1/
for all n > 0.
(2.11)
“mcs” — 2017/6/5 — 19:42 — page 44 — #52
44
Chapter 2
The Well Ordering Principle
Problem 2.17.
Prove by the Well Ordering Principle that for all nonnegative integers, n:
3
3
3
3
0 C 1 C 2 C C n D
n.n C 1/
2
2
:
(2.12)
Problem 2.18.
Use the Well Ordering Principle to prove that
1 2 C 2 3 C 3 4 C C n.n C 1/ D
n.n C 1/.n C 2/
3
(*)
for all integers n 1.
Problem 2.19.
Say a number of cents is makeable if it is the value of some set of 6 cent and 15
cent stamps. Use the Well Ordering Principle to show that every integer that is a
multiple of 3 and greater than or equal to twelve is makeable.
Problems for Section 2.4
Homework Problems
Problem 2.20.
Complete the proof of Lemma 2.4.6 by showing that the number nS C fS is the
minimum element in S .
Practice Problems
Problem 2.21.
Indicate which of the following sets of numbers have a minimum element and
which are well ordered. For those that are not well ordered, give an example of
a subset with no minimum element.
p
(a) The integers 2.
p
(b) The rational numbers 2.
(c) The set of rationals of the form 1=n where n is a positive integer.
“mcs” — 2017/6/5 — 19:42 — page 45 — #53
2.4. Well Ordered Sets
45
(d) The set G of rationals of the form m=n where m; n > 0 and n g, where g
is a googol 10100 .
(e) The set F of fractions of the form n=.n C 1/:
0 1 2 3
; ; ; ;::::
1 2 3 4
(f) Let W WWD N [ F be the set consisting of the nonnegative integers along with
all the fractions of the form n=.n C 1/. Describe a length 5 decreasing sequence of
elements of W starting with 1,. . . length 50 decreasing sequence,. . . length 500.
Problem 2.22.
Use the Well Ordering Principle to prove that every finite, nonempty set of real
numbers has a minimum element.
Class Problems
Problem 2.23.
Prove that a set R of real numbers is well ordered iff there is no infinite decreasing
sequence of numbers R. In other words, there is no set of numbers ri 2 R such
that
r0 > r1 > r2 > : : : :
(2.13)
“mcs” — 2017/6/5 — 19:42 — page 46 — #54
“mcs” — 2017/6/5 — 19:42 — page 47 — #55
3
Logical Formulas
It is amazing that people manage to cope with all the ambiguities in the English
language. Here are some sentences that illustrate the issue:
“You may have cake, or you may have ice cream.”
“If pigs can fly, then your account won’t get hacked.”
“If you can solve any problem we come up with, then you get an A for the
course.”
“Every American has a dream.”
What precisely do these sentences mean? Can you have both cake and ice cream or
must you choose just one dessert? Pigs can’t fly, so does the second sentence say
anything about the security of your account? If you can solve some problems we
come up with, can you get an A for the course? And if you can’t solve a single one
of the problems, does it mean you can’t get an A? Finally, does the last sentence
imply that all Americans have the same dream—say of owning a house—or might
different Americans have different dreams—say, Eric dreams of designing a killer
software application, Tom of being a tennis champion, Albert of being able to sing?
Some uncertainty is tolerable in normal conversation. But when we need to
formulate ideas precisely—as in mathematics and programming—the ambiguities
inherent in everyday language can be a real problem. We can’t hope to make an
exact argument if we’re not sure exactly what the statements mean. So before we
start into mathematics, we need to investigate the problem of how to talk about
mathematics.
To get around the ambiguity of English, mathematicians have devised a special language for talking about logical relationships. This language mostly uses
ordinary English words and phrases such as “or,” “implies,” and “for all.” But
mathematicians give these words precise and unambiguous definitions which don’t
always match common usage.
Surprisingly, in the midst of learning the language of logic, we’ll come across
the most important open problem in computer science—a problem whose solution
could change the world.
“mcs” — 2017/6/5 — 19:42 — page 48 — #56
48
3.1
Chapter 3
Logical Formulas
Propositions from Propositions
In English, we can modify, combine, and relate propositions with words such as
“not,” “and,” “or,” “implies,” and “if-then.” For example, we can combine three
propositions into one like this:
If all humans are mortal and all Greeks are human, then all Greeks are mortal.
For the next while, we won’t be much concerned with the internals of propositions—
whether they involve mathematics or Greek mortality—but rather with how propositions are combined and related. So, we’ll frequently use variables such as P and
Q in place of specific propositions such as “All humans are mortal” and “2 C 3 D
5.” The understanding is that these propositional variables, like propositions, can
take on only the values T (true) and F (false). Propositional variables are also
called Boolean variables after their inventor, the nineteenth century mathematician
George—you guessed it—Boole.
3.1.1
NOT , AND ,
and OR
Mathematicians use the words NOT, AND and OR for operations that change or
combine propositions. The precise mathematical meaning of these special words
can be specified by truth tables. For example, if P is a proposition, then so is
“NOT.P /,” and the truth value of the proposition “NOT.P /” is determined by the
truth value of P according to the following truth table:
P NOT.P /
T
F
F
T
The first row of the table indicates that when proposition P is true, the proposition “NOT.P /” is false. The second line indicates that when P is false, “NOT.P /”
is true. This is probably what you would expect.
In general, a truth table indicates the true/false value of a proposition for each
possible set of truth values for the variables. For example, the truth table for the
proposition “P AND Q” has four lines, since there are four settings of truth values
for the two variables:
P Q P AND Q
T T
T
T F
F
F T
F
F F
F
“mcs” — 2017/6/5 — 19:42 — page 49 — #57
3.1. Propositions from Propositions
49
According to this table, the proposition “P AND Q” is true only when P and Q
are both true. This is probably the way you ordinarily think about the word “and.”
There is a subtlety in the truth table for “P OR Q”:
P Q P OR Q
T T
T
T F
T
F T
T
F F
F
The first row of this table says that “P OR Q” is true even if both P and Q are
true. This isn’t always the intended meaning of “or” in everyday speech, but this is
the standard definition in mathematical writing. So if a mathematician says, “You
may have cake, or you may have ice cream,” he means that you could have both.
If you want to exclude the possibility of having both cake and ice cream, you
should combine them with the exclusive-or operation, XOR:
P Q P XOR Q
T T
F
T F
T
F T
T
F F
F
3.1.2
If and Only If
Mathematicians commonly join propositions in an additional way that doesn’t arise
in ordinary speech. The proposition “P if and only if Q” asserts that P and Q have
the same truth value. Either both are true or both are false.
P Q P IFF Q
T T
T
T F
F
F T
F
F F
T
For example, the following if-and-only-if statement is true for every real number
x:
x 2 4 0 IFF jxj 2:
For some values of x, both inequalities are true. For other values of x, neither
inequality is true. In every case, however, the IFF proposition as a whole is true.
“mcs” — 2017/6/5 — 19:42 — page 50 — #58
50
Chapter 3
3.1.3
Logical Formulas
IMPLIES
The combining operation whose technical meaning is least intuitive is “implies.”
Here is its truth table, with the lines labeled so we can refer to them later.
P Q P IMPLIES Q
T T
T
F
T F
F T
T
F F
T
(tt)
(tf)
(ft)
(ff)
The truth table for implications can be summarized in words as follows:
An implication is true exactly when the if-part is false or the then-part is true.
This sentence is worth remembering; a large fraction of all mathematical statements
are of the if-then form!
Let’s experiment with this definition. For example, is the following proposition
true or false?
If Goldbach’s Conjecture is true, then x 2 0 for every real number x.
We already mentioned that no one knows whether Goldbach’s Conjecture, Proposition 1.1.6, is true or false. But that doesn’t prevent us from answering the question!
This proposition has the form P IMPLIES Q where the hypothesis P is “Goldbach’s Conjecture is true” and the conclusion Q is “x 2 0 for every real number
x.” Since the conclusion is definitely true, we’re on either line (tt) or line (ft) of the
truth table. Either way, the proposition as a whole is true!
Now let’s figure out the truth of one of our original examples:
If pigs fly, then your account won’t get hacked.
Forget about pigs, we just need to figure out whether this proposition is true or
false. Pigs do not fly, so we’re on either line (ft) or line (ff) of the truth table. In
both cases, the proposition is true!
False Hypotheses
This mathematical convention—that an implication as a whole is considered true
when its hypothesis is false—contrasts with common cases where implications are
supposed to have some causal connection between their hypotheses and conclusions.
For example, we could agree—or at least hope—that the following statement is
true:
“mcs” — 2017/6/5 — 19:42 — page 51 — #59
3.1. Propositions from Propositions
51
If you followed the security protocal, then your account won’t get
hacked.
We regard this implication as unproblematical because of the clear causal connection between security protocols and account hackability.
On the other hand, the statement:
If pigs could fly, then your account won’t get hacked,
would commonly be rejected as false—or at least silly—because porcine aeronautics have nothing to do with your account security. But mathematically, this implication counts as true.
It’s important to accept the fact that mathematical implications ignore causal
connections. This makes them a lot simpler than causal implications, but useful
nevertheless. To illustrate this, suppose we have a system specification which consists of a series of, say, a dozen rules,1
If
If
If
the
then
the
then
system sensors are in condition 1,
the system takes action 1.
system sensors are in condition 2,
the system takes action 2.
::
:
the
then
system sensors are in condition 12,
the system takes action 12.
Letting Ci be the proposition that the system sensors are in condition i , and Ai
be the proposition that system takes action i , the specification can be restated more
concisely by the logical formulas
C1 IMPLIES A1 ;
C2 IMPLIES A2 ;
::
:
C12 IMPLIES A12 :
Now the proposition that the system obeys the specification can be nicely expressed
as a single logical formula by combining the formulas together with ANDs::
ŒC1 IMPLIES A1  AND ŒC2 IMPLIES A2  AND AND ŒC12 IMPLIES A12 : (3.1)
For example, suppose only conditions C2 and C5 are true, and the system indeed
takes the specified actions A2 and A5 . So in this case, the system is behaving
1
Problem 3.16 concerns just such a system.
“mcs” — 2017/6/5 — 19:42 — page 52 — #60
52
Chapter 3
Logical Formulas
according to specification, and we accordingly want formula (3.1) to come out true.
The implications C2 IMPLIES A2 and C5 IMPLIES A5 are both true because both
their hypotheses and their conclusions are true. But in order for (3.1) to be true, we
need all the other implications, all of whose hypotheses are false, to be true. This
is exactly what the rule for mathematical implications accomplishes.
3.2
Propositional Logic in Computer Programs
Propositions and logical connectives arise all the time in computer programs. For
example, consider the following snippet, which could be either C, C++, or Java:
if ( x > 0 || (x <= 0 && y > 100) )
::
:
(further instructions)
Java uses the symbol || for “OR,” and the symbol && for “AND.” The further
instructions are carried out only if the proposition following the word if is true.
On closer inspection, this big expression is built from two simpler propositions.
Let A be the proposition that x > 0, and let B be the proposition that y > 100.
Then we can rewrite the condition as
A OR .NOT.A/ AND B/:
3.2.1
(3.2)
Truth Table Calculation
A truth table calculation reveals that the more complicated expression 3.2 always
has the same truth value as
A OR B:
(3.3)
We begin with a table with just the truth values of A and B:
A B A OR .NOT.A/ AND B/ A OR B
T T
T F
F T
F F
“mcs” — 2017/6/5 — 19:42 — page 53 — #61
3.2. Propositional Logic in Computer Programs
53
These values are enough to fill in two more columns:
A B A OR .NOT.A/ AND B/ A OR B
T T
F
T
F
T
T F
F T
T
T
T
F
F F
Now we have the values needed to fill in the AND column:
A B A OR .NOT.A/ AND B/ A OR B
T T
F
F
T
T F
F
F
T
T
T
T
F T
F F
T
F
F
and this provides the values needed to fill in the remaining column for the first OR:
A B A OR .NOT.A/ AND B/ A OR B
T T
T
F
F
T
T F
T
F
F
T
T
T
T
T
F T
F F
F
T
F
F
Expressions whose truth values always match are called equivalent. Since the two
emphasized columns of truth values of the two expressions are the same, they are
equivalent. So we can simplify the code snippet without changing the program’s
behavior by replacing the complicated expression with an equivalent simpler one:
if ( x > 0 || y > 100 )
::
:
(further instructions)
The equivalence of (3.2) and (3.3) can also be confirmed reasoning by cases:
A is T. An expression of the form .T OR anything/ is equivalent to T. Since A is T
both (3.2) and (3.3) in this case are of this form, so they have the same truth
value, namely, T.
A is F. An expression of the form .F OR anything/ will have same truth value as
anything. Since A is F, (3.3) has the same truth value as B.
An expression of the form .T AND anything/ is equivalent to anything, as is
any expression of the form F OR anything. So in this case A OR .NOT.A/ AND
B/ is equivalent to .NOT.A/ AND B/, which in turn is equivalent to B.
“mcs” — 2017/6/5 — 19:42 — page 54 — #62
54
Chapter 3
Logical Formulas
Therefore both (3.2) and (3.3) will have the same truth value in this case,
namely, the value of B.
Simplifying logical expressions has real practical importance in computer science. Expression simplification in programs like the one above can make a program
easier to read and understand. Simplified programs may also run faster, since they
require fewer operations. In hardware, simplifying expressions can decrease the
number of logic gates on a chip because digital circuits can be described by logical
formulas (see Problems 3.6 and 3.7). Minimizing the logical formulas corresponds
to reducing the number of gates in the circuit. The payoff of gate minimization is
potentially enormous: a chip with fewer gates is smaller, consumes less power, has
a lower defect rate, and is cheaper to manufacture.
3.2.2
Cryptic Notation
Java uses symbols like “&&” and “jj” in place of AND and OR. Circuit designers
use “” and “C,” and actually refer to AND as a product and OR as a sum. Mathematicians use still other symbols, given in the table below.
English
Symbolic Notation
NOT .P /
P AND Q
P OR Q
P IMPLIES Q
if P then Q
P IFF Q
P XOR Q
:P (alternatively, P )
P ^Q
P _Q
P !Q
P !Q
P !Q
P ˚Q
For example, using this notation, “If P AND NOT.Q/, then R” would be written:
.P ^ Q/ ! R:
The mathematical notation is concise but cryptic. Words such as “AND” and
“OR” are easier to remember and won’t get confused with operations on numbers.
We will often use P as an abbreviation for NOT.P /, but aside from that, we mostly
stick to the words—except when formulas would otherwise run off the page.
3.3
Equivalence and Validity
3.3.1
Implications and Contrapositives
Do these two sentences say the same thing?
“mcs” — 2017/6/5 — 19:42 — page 55 — #63
3.3. Equivalence and Validity
55
If I am hungry, then I am grumpy.
If I am not grumpy, then I am not hungry.
We can settle the issue by recasting both sentences in terms of propositional logic.
Let P be the proposition “I am hungry” and Q be “I am grumpy.” The first sentence
says “P IMPLIES Q” and the second says “NOT.Q/ IMPLIES NOT.P /.” Once
more, we can compare these two statements in a truth table:
P Q .P IMPLIES Q/ .NOT.Q/ IMPLIES NOT.P //
T T
T
F
T
F
T F
F
T
F
F
F T
T
F
T
T
T
T
T
T
F F
Sure enough, the highlighted columns showing the truth values of these two statements are the same. A statement of the form “NOT.Q/ IMPLIES NOT.P /” is called
the contrapositive of the implication “P IMPLIES Q.” The truth table shows that
an implication and its contrapositive are equivalent—they are just different ways of
saying the same thing.
In contrast, the converse of “P IMPLIES Q” is the statement “Q IMPLIES P .”
The converse to our example is:
If I am grumpy, then I am hungry.
This sounds like a rather different contention, and a truth table confirms this suspicion:
P Q P IMPLIES Q Q IMPLIES P
T T
T
T
F
T
T F
F T
T
F
F F
T
T
Now the highlighted columns differ in the second and third row, confirming that an
implication is generally not equivalent to its converse.
One final relationship: an implication and its converse together are equivalent to
an iff statement, specifically, to these two statements together. For example,
If I am grumpy then I am hungry, and if I am hungry then I am grumpy.
are equivalent to the single statement:
I am grumpy iff I am hungry.
“mcs” — 2017/6/5 — 19:42 — page 56 — #64
56
Chapter 3
Logical Formulas
Once again, we can verify this with a truth table.
P Q .P IMPLIES Q/ AND .Q IMPLIES P / P IFF Q
T T
T
T
T
T
T F
F
F
T
F
F T
T
F
F
F
T
T
T
T
F F
The fourth column giving the truth values of
.P IMPLIES Q/ AND .Q IMPLIES P /
is the same as the sixth column giving the truth values of P IFF Q, which confirms
that the AND of the implications is equivalent to the IFF statement.
3.3.2
Validity and Satisfiability
A valid formula is one which is always true, no matter what truth values its variables may have. The simplest example is
P OR NOT.P /:
You can think about valid formulas as capturing fundamental logical truths. For
example, a property of implication that we take for granted is that if one statement
implies a second one, and the second one implies a third, then the first implies the
third. The following valid formula confirms the truth of this property of implication.
Œ.P IMPLIES Q/ AND .Q IMPLIES R/ IMPLIES .P IMPLIES R/:
Equivalence of formulas is really a special case of validity. Namely, statements
F and G are equivalent precisely when the statement .F IFF G/ is valid. For
example, the equivalence of the expressions (3.3) and (3.2) means that
.A OR B/ IFF .A OR .NOT.A/ AND B//
is valid. Of course, validity can also be viewed as an aspect of equivalence. Namely,
a formula is valid iff it is equivalent to T.
A satisfiable formula is one which can sometimes be true—that is, there is some
assignment of truth values to its variables that makes it true. One way satisfiability comes up is when there are a collection of system specifications. The job of
the system designer is to come up with a system that follows all the specs. This
means that the AND of all the specs must be satisfiable or the designer’s job will be
impossible (see Problem 3.16).
There is also a close relationship between validity and satisfiability: a statement
P is satisfiable iff its negation NOT.P / is not valid.
“mcs” — 2017/6/5 — 19:42 — page 57 — #65
3.4. The Algebra of Propositions
3.4
57
The Algebra of Propositions
3.4.1
Propositions in Normal Form
Every propositional formula is equivalent to a “sum-of-products” or disjunctive
normal form (DNF).
More precisely, a propositional variable A or its negation bar.A/ is called a
literal, and an AND of literals involving distinct variables is called an AND-clause.
For example,
A AND B AND C
is an AND-clause, but A AND B AND B AND C is not because B appears twice.
Finally, a DNF is an ORof AND-clauses such as
.A AND B/ OR .A AND C /:
(3.4)
You can read a DNF for any propositional formula directly from its truth table.
For example, the formula
A AND .B OR C /
(3.5)
has truth table:
A
T
T
T
T
F
F
F
F
B C A AND .B OR C /
T T
T
T F
T
F T
T
F F
F
T T
F
T F
F
F T
F
F F
F
The formula (3.5) is true in the first row when A, B and C are all true, that is, where
A AND B AND C is true. It is also true in the second row where A AND B AND C
is true, and in the third row when A AND B AND C is true, and that’s all. So (3.5)
is true exactly when
.A AND B AND C / OR .A AND B AND C / OR .A AND B AND C /
(3.6)
is true.
The expression (3.6) is a DNF where each AND-clause actually includes a literal
for every one of the variables in the whole formula. We’ll call such a formula a full
DNF.
“mcs” — 2017/6/5 — 19:42 — page 58 — #66
58
Chapter 3
Logical Formulas
A DNF formula can often be simplified into a smaller DNF For example, the
DNF (3.6) further simplifies to the equivalent DNF (3.4) above.
Applying the same reasoning to the F entries of a truth table yields a conjunctive
normal form (CNF) for any formula—an AND of OR-clauses, where an OR-clause
is an OR of literals from different variables.
For example, formula (3.5) is false in the fourth row of its truth table (3.4.1)
where A is T, B is F and C is F. But this is exactly the one row where the ORclause .A OR B OR C / is F! Likewise, (3.5) is false in the fifth row, which is
exactly where .A OR B OR C / is F. This means that (3.5) will be F whenever the
AND of these two OR -clauses is false. Continuing in this way with the OR -clauses
corresponding to the remaining three rows where (3.5) is false, we get a CNF that
is equivalent to (3.5), namely,
.A OR B OR C / AND .A OR B OR C / AND .A OR B OR C /AND
.A OR B OR C / AND .A OR B OR C /
Again, each OR-clause includes a literal for every one of the variables, that is, it is
a full CNF
The methods above can be applied to any truth table, which implies
Theorem 3.4.1. Every propositional formula is equivalent to both a full disjunctive
normal form and a full conjunctive normal form.
3.4.2
Proving Equivalences
A check of equivalence or validity by truth table runs out of steam pretty quickly:
a proposition with n variables has a truth table with 2n lines, so the effort required
to check a proposition grows exponentially with the number of variables. For a
proposition with just 30 variables, that’s already over a billion lines to check!
An alternative approach that sometimes helps is to use algebra to prove equivalence. A lot of different operators may appear in a propositional formula, so a
useful first step is to get rid of all but three: AND, OR and NOT. This is easy because each of the operators is equivalent to a simple formula using only these three.
For example, A IMPLIES B is equivalent to NOT.A/ OR B. Formulas defining the
remaining operators using only QAND; OR and NOT are left to Problem 3.17.
We list below a bunch of equivalence axioms with the symbol “ ! ” between
equivalent formulas. These axioms are important because they are all that’s needed
to prove every possible equivalence. We’ll start with some equivalences for AND’s
“mcs” — 2017/6/5 — 19:42 — page 59 — #67
3.4. The Algebra of Propositions
59
that look like the familiar ones for multiplication of numbers:
! B AND A
A AND B
(commutativity of AND)
(3.7)
! A AND .B AND C /
.A AND B/ AND C
(associativity of AND)
(3.8)
T AND A
! A
(identity for AND)
F AND A
! F
(zero for AND)
! .A AND B/ OR .A AND C / (distributivity of AND over OR)
(3.9)
A AND .B OR C /
Associativity (3.8) justifies writing A AND B AND C without specifying whether
it is parenthesized as A AND .B AND C / or .A AND B/ AND C . Both ways of
inserting parentheses yield equivalent formulas.
Unlike arithmetic rules for numbers, there is also a distributivity law for “sums”
over “products:”
A OR .B AND C /
! .A OR B/ AND .A OR C / (distributivity of OR over AND)
(3.10)
Three more axioms that don’t directly correspond to number properties are
A AND A
! A
(idempotence for AND)
A AND A
! F
(contradiction for AND)
(3.11)
NOT .A/
! A
(double negation)
(3.12)
There are a corresponding set of equivalences for OR which we won’t bother to
list, except for validity rule (3.13) for OR:
A OR A
! T
(validity for OR)
(3.13)
Finally, there are De Morgan’s Laws which explain how to distribute NOT’s over
AND ’s and OR ’s:
NOT.A AND
B/
! A OR B
NOT .A OR
B/
! A AND B
(De Morgan for AND)
(3.14)
(De Morgan for OR)
(3.15)
All of these axioms can be verified easily with truth tables.
These axioms are all that’s needed to convert any formula to a full DNF. We can
illustrate how they work by applying them to turn the negation of formula (3.5),
NOT..A AND
B/ OR .A AND C //:
(3.16)
“mcs” — 2017/6/5 — 19:42 — page 60 — #68
60
Chapter 3
Logical Formulas
into a full DNF.
We start by applying De Morgan’s Law for OR (3.15) to (3.16) in order to move
the NOT deeper into the formula. This gives
NOT .A AND
B/ AND NOT.A AND C /:
Now applying De Morgan’s Law for AND (3.14) to the two innermost AND-terms,
gives
.A OR B/ AND .A OR C /:
(3.17)
At this point NOT only applies to variables, and we won’t need De Morgan’s Laws
any further.
Now we will repeatedly apply (3.9), distributivity of AND over OR, to turn (3.17)
into a DNF. To start, we’ll distribute .A OR B/ over AND to get
..A OR B/ AND A/ OR ..A OR B/ AND C /:
Using distributivity over both AND’s we get
..A AND A/ OR .B AND A// OR ..A AND C / OR .B AND C //:
By the way, we’ve implicitly used commutativity (3.7) here to justify distributing
over an AND from the right. Now applying idempotence to remove the duplicate
occurrence of A we get
.A OR .B AND A// OR ..A AND C / OR .B AND C //:
Associativity of QOR now allows dropping the parentheses grouping the ANDclauses to yield the following DNF for (3.16):
A OR .B AND A/ OR .A AND C / OR .B AND C /:
(3.18)
The penultimate step is to turn this DNF into a full DNF. This can be done
separately for ech AND-clause. We’ll illustrate how using the second AND-clause
.B AND A/. This clause needs to mention C to be in full form. To introduce C , we
use validity for OR and identity for AND to conclude that
.B AND A/
! .B AND A/ AND .C OR C /:
Now distributing .B AND A/ over the OR in .C OR C / yields the full DNF
.B AND A AND C / OR .B AND A AND C /:
“mcs” — 2017/6/5 — 19:42 — page 61 — #69
3.4. The Algebra of Propositions
61
Doing the same thing to the other AND-clauses in (3.18) finally gives a full DNF
for (3.5):
.A AND B AND C / OR .A AND B AND C / OR
.A AND B AND C / OR .A AND B AND C / OR
.B AND A AND C / OR .B AND A AND C / OR
.A AND C AND B/ OR .A AND C AND B/ OR
.B AND C AND A/ OR .B AND C AND A/:
The final step is to use commutativity to sort the variables within the AND-clauses
and then sort the AND-clauses themselves, followed by applying OR-idempotence
as needed to remove duplicate AND-clauses. This finally yields a sorted full DNF
without duplicates which is called a canonical DNF :
.A AND B
.A AND B
.A AND B
.A AND B
.A AND B
C / OR
C / OR
AND C / OR
AND C / OR
AND C /:
AND
AND
This example illustrates the general strategy for applying the axioms above to any
given propositional formala to derive an equivalent canonical DNF. This proves:
Theorem 3.4.2. Using the equivalences listed above, any propositional formula
can be proved equivalent to a canonical form.
What has this got to do with equivalence? That’s easy: to prove that two formulas
are equivalent, convert them both to canonical forms over the set of variables that
appear in at least one of the formulas—call these the combined variables. Now
if two formulas are equivalent to the same canonical form then the formula are
certainly equivalent. Conversely, the way we read off a full disjunctive normal
form from a truth table actually yields a canonical form. So if two formulas are
equivalent, they will have the same truth table over the combined variables, and
therefore they will have the same canonical form. This proves
Theorem 3.4.3 (Completeness of the propositional equivalence axioms). Two propositional formula are equivalent iff they can be proved equivalent using the equivalence axioms listed above.
Notice that the same approach could be taken used CNF instead of DNF canonical forms.
The benefit of the axioms is that they allow some ingenious proofs of equivalence that may involve much less effort than the truth table method. Moreover,
“mcs” — 2017/6/5 — 19:42 — page 62 — #70
62
Chapter 3
Logical Formulas
Theorem 3.4.3 reassures us that the axioms are guaranteed to provide a proof of
every equivalence, which is a great punchline for this section.
But we don’t want to mislead you: the guaranteed proof involves deriving canonical forms, and canonical forms are essentially copies of truth tables. There is no
reason to expect algebraic proofs of equivalence to be any easier in general than
conversion to canonical form, which means algebraic proofs will generally be no
easier than using truth tables.
3.5
The SAT Problem
Determining whether or not a more complicated proposition is satisfiable is not so
easy. How about this one?
.P OR Q OR R/ AND .P OR Q/ AND .P OR R/ AND .R OR Q/
The general problem of deciding whether a proposition is satisfiable is called
SAT. One approach to SAT is to construct a truth table and check whether or not
a T ever appears, but as with testing validity, this approach quickly bogs down
for formulas with many variables because truth tables grow exponentially with the
number of variables.
Is there a more efficient solution to SAT? In particular, is there some brilliant
procedure that determines SAT in a number of steps that grows polynomially—like
n2 or n14 —instead of exponentially—2n —whether any given proposition of size n
is satisfiable or not? No one knows. And an awful lot hangs on the answer.
The general definition of an “efficient” procedure is one that runs in polynomial
time, that is, that runs in a number of basic steps bounded by a polynomial in s,
where s is the size of an input. It turns out that an efficient solution to SAT would
immediately imply efficient solutions to many other important problems involving
scheduling, routing, resource allocation, and circuit verification across multiple disciplines including programming, algebra, finance, and political theory. This would
be wonderful, but there would also be worldwide chaos. Decrypting coded messages would also become an easy task, so online financial transactions would be
insecure and secret communications could be read by everyone. Why this would
happen is explained in Section 9.12.
Of course, the situation is the same for validity checking, since you can check for
validity by checking for satisfiability of a negated formula. This also explains why
the simplification of formulas mentioned in Section 3.2 would be hard—validity
testing is a special case of determining if a formula simplifies to T.
“mcs” — 2017/6/5 — 19:42 — page 63 — #71
3.6. Predicate Formulas
63
Recently there has been exciting progress on SAT-solvers for practical applications like digital circuit verification. These programs find satisfying assignments
with amazing efficiency even for formulas with millions of variables. Unfortunately, it’s hard to predict which kind of formulas are amenable to SAT-solver
methods, and for formulas that are unsatisfiable, SAT-solvers are generally much
less effective.
So no one has a good idea how to solve SAT in polynomial time, or how to
prove that it can’t be done—researchers are completely stuck. The problem of
determining whether or not SAT has a polynomial time solution is known as the
“P vs. NP” problem.2 It is the outstanding unanswered question in theoretical
computer science. It is also one of the seven Millenium Problems: the Clay Institute
will award you $1,000,000 if you solve the P vs. NP problem.
3.6
Predicate Formulas
3.6.1
Quantifiers
The “for all” notation 8 has already made an early appearance in Section 1.1. For
example, the predicate
“x 2 0”
is always true when x is a real number. That is,
8x 2 R: x 2 0
is a true statement. On the other hand, the predicate
“5x 2
7 D 0”
p
is only sometimes true; specifically, when x D ˙ 7=5. There is a “there exists”
notation 9 to indicate that a predicate is true for at least one, but not necessarily all
objects. So
9x 2 R: 5x 2 7 D 0
is true, while
8x 2 R: 5x 2
7D0
is not true.
2 P stands for problems whose instances can be solved in time that grows polynomially with the
size of the instance. NP stands for nondeterministtic polynomial time, but we’ll leave an explanation
of what that is to texts on the theory of computational complexity.
“mcs” — 2017/6/5 — 19:42 — page 64 — #72
64
Chapter 3
Logical Formulas
There are several ways to express the notions of “always true” and “sometimes
true” in English. The table below gives some general formats on the left and specific
examples using those formats on the right. You can expect to see such phrases
hundreds of times in mathematical writing!
Always True
For all x 2 D, P .x/ is true.
P .x/ is true for every x in the set D.
For all x 2 R, x 2 0.
x 2 0 for every x 2 R.
Sometimes True
There is an x 2 D such that P .x/ is true.
P .x/ is true for some x in the set D.
P .x/ is true for at least one x 2 D.
There is an x 2 R such that 5x 2 7 D 0.
5x 2 7 D 0 for some x 2 R.
5x 2 7 D 0 for at least one x 2 R.
All these sentences “quantify” how often the predicate is true. Specifically, an
assertion that a predicate is always true is called a universal quantification, and an
assertion that a predicate is sometimes true is an existential quantification. Sometimes the English sentences are unclear with respect to quantification:
If you can solve any problem we come up with,
then you get an A for the course.
(3.19)
The phrase “you can solve any problem we can come up with” could reasonably be
interpreted as either a universal or existential quantification:
you can solve every problem we come up with,
(3.20)
you can solve at least one problem we come up with.
(3.21)
or maybe
To be precise, let Probs be the set of problems we come up with, Solves.x/ be
the predicate “You can solve problem x,” and G be the proposition, “You get an A
for the course.” Then the two different interpretations of (3.19) can be written as
follows:
3.6.2
.8x 2 Probs: Solves.x// IMPLIES G;
for (3.20);
.9x 2 Probs: Solves.x// IMPLIES G:
for (3.21):
Mixing Quantifiers
Many mathematical statements involve several quantifiers. For example, we already described
“mcs” — 2017/6/5 — 19:42 — page 65 — #73
3.6. Predicate Formulas
65
Goldbach’s Conjecture 1.1.6: Every even integer greater than 2 is the
sum of two primes.
Let’s write this out in more detail to be precise about the quantification:
For every even integer n greater than 2, there exist primes p and q such
that n D p C q.
Let Evens be the set of even integers greater than 2, and let Primes be the set of
primes. Then we can write Goldbach’s Conjecture in logic notation as follows:
8n
Evens
2 Primes 9q 2 Primes: n D p C q:
„ 2ƒ‚
… 9p
ƒ‚
…
„
for every even
integer n > 2
3.6.3
there exist primes
p and q such that
Order of Quantifiers
Swapping the order of different kinds of quantifiers (existential or universal) usually
changes the meaning of a proposition. For example, let’s return to one of our initial,
confusing statements:
“Every American has a dream.”
This sentence is ambiguous because the order of quantifiers is unclear. Let A be
the set of Americans, let D be the set of dreams, and define the predicate H.a; d /
to be “American a has dream d .” Now the sentence could mean there is a single
dream that every American shares—such as the dream of owning their own home:
9 d 2 D 8a 2 A: H.a; d /
Or it could mean that every American has a personal dream:
8a 2 A 9 d 2 D: H.a; d /
For example, some Americans may dream of a peaceful retirement, while others
dream of continuing practicing their profession as long as they live, and still others
may dream of being so rich they needn’t think about work at all.
Swapping quantifiers in Goldbach’s Conjecture creates a patently false statement
that every even number 2 is the sum of the same two primes:
9 p 2 Primes 9 q 2 Primes: 8n
2 Evens n D p C q:
„
ƒ‚
… „ ƒ‚ …
there exist primes
p and q such that
for every even
integer n > 2
“mcs” — 2017/6/5 — 19:42 — page 66 — #74
66
Chapter 3
3.6.4
Logical Formulas
Variables Over One Domain
When all the variables in a formula are understood to take values from the same
nonempty set D it’s conventional to omit mention of D. For example, instead of
8x 2 D 9y 2 D: Q.x; y/ we’d write 8x9y: Q.x; y/. The unnamed nonempty set
that x and y range over is called the domain of discourse, or just plain domain, of
the formula.
It’s easy to arrange for all the variables to range over one domain. For example, Goldbach’s Conjecture could be expressed with all variables ranging over the
domain N as
8n: n 2 Evens IMPLIES .9 p 9 q: p 2 Primes AND q 2 Primes AND n D p C q/:
3.6.5
Negating Quantifiers
There is a simple relationship between the two kinds of quantifiers. The following
two sentences mean the same thing:
Not everyone likes ice cream.
There is someone who does not like ice cream.
The equivalence of these sentences is an instance of a general equivalence that
holds between predicate formulas:
NOT .8x: P .x//
is equivalent to 9x: NOT.P .x//:
(3.22)
Similarly, these sentences mean the same thing:
There is no one who likes being mocked.
Everyone dislikes being mocked.
The corresponding predicate formula equivalence is
NOT .9x: P .x//
is equivalent to 8x: NOT.P .x//:
(3.23)
Note that the equivalence (3.23) follows directly by negating both sides the equivalence (3.22).
The general principle is that moving a NOT to the other side of an “9” changes
it into “8,” and vice versa.
These equivalences are called De Morgan’s Laws for Quantifiers because they
can be understood as applying De Morgan’s Laws for propositional formulas to
an infinite sequence of AND’s and OR’s. For example, we can explain (3.22) by
“mcs” — 2017/6/5 — 19:42 — page 67 — #75
3.6. Predicate Formulas
67
supposing the domain of discourse is fd0 ; d1 ; : : : ; dn ; : : :g. Then 9x: NOT.P .x//
means the same thing as the infinite OR:
NOT.P .d0 // OR NOT .P .d1 // OR
OR NOT.P .dn // OR : : : :
(3.24)
Applying De Morgan’s rule to this infinite OR yields the equivalent formula
NOT ŒP .d0 / AND
P .d1 / AND AND P .dn / AND : : : :
(3.25)
But (3.25) means the same thing as
NOTŒ8x: P .x/:
This explains why 9x: NOT.P .x// means the same thing as NOTŒ8x: P .x/, which
confirms(3.22).
3.6.6
Validity for Predicate Formulas
The idea of validity extends to predicate formulas, but to be valid, a formula now
must evaluate to true no matter what the domain of discourse may be, no matter
what values its variables may take over the domain, and no matter what interpretations its predicate variables may be given. For example, the equivalence (3.22) that
gives the rule for negating a universal quantifier means that the following formula
is valid:
NOT.8x: P .x// IFF 9x: NOT .P .x//:
(3.26)
Another useful example of a valid assertion is
9x8y: P .x; y/ IMPLIES 8y9x: P .x; y/:
(3.27)
Here’s an explanation why this is valid:
Let D be the domain for the variables and P0 be some binary predicate3 on D. We need to show that if
9x 2 D: 8y 2 D: P0 .x; y/
(3.28)
holds under this interpretation, then so does
8y 2 D 9x 2 D: P0 .x; y/:
(3.29)
So suppose (3.28) is true. Then by definition of 9, this means that some
element d0 2 D has the property that
8y 2 D: P0 .d0 ; y/:
3 That
is, a predicate that depends on two variables.
“mcs” — 2017/6/5 — 19:42 — page 68 — #76
68
Chapter 3
Logical Formulas
By definition of 8, this means that
P0 .d0 ; d /
is true for all d 2 D. So given any d 2 D, there is an element in D,
namely d0 , such that P0 .d0 ; d / is true. But that’s exactly what (3.29)
means, so we’ve proved that (3.29) holds under this interpretation, as
required.
We hope this is helpful as an explanation, but we don’t really want to call it a
“proof.” The problem is that with something as basic as (3.27), it’s hard to see
what more elementary axioms are ok to use in proving it. What the explanation
above did was translate the logical formula (3.27) into English and then appeal to
the meaning, in English, of “for all” and “there exists” as justification.
In contrast to (3.27), the formula
8y9x: P .x; y/ IMPLIES 9x8y: P .x; y/:
(3.30)
is not valid. We can prove this just by describing an interpretation where the hypothesis 8y9x: P .x; y/ is true but the conclusion 9x8y: P .x; y/ is not true. For
example, let the domain be the integers and P .x; y/ mean x > y. Then the hypothesis would be true because, given a value n for y we could choose the value
of x to be n C 1, for example. But under this interpretation the conclusion asserts
that there is an integer that is bigger than all integers, which is certainly false. An
interpretation like this that falsifies an assertion is called a counter-model to that
assertion.
3.7
References
[21]
Problems for Section 3.1
Practice Problems
Problem 3.1.
Some people are uncomfortable with the idea that from a false hypothesis you can
prove everything, and instead of having P IMPLIES Q be true when P is false,
“mcs” — 2017/6/5 — 19:42 — page 69 — #77
3.7. References
69
they want P IMPLIES Q to be false when P is false. This would lead to IMPLIES
having the same truth table as what propositional connective?
Problem 3.2.
Your class has a textbook and a final exam. Let P , Q and R be the following
propositions:
P WWD You get an A on the final exam.
QWWD You do every exercise in the book.
RWWD You get an A in the class.
Translate following assertions into propositional formulas using P , Q, R and
the propositional connectives AND; NOT; IMPLIES.
(a) You get an A in the class, but you do not do every exercise in the book.
(b) You get an A on the final, you do every exercise in the book, and you get an A
in the class.
(c) To get an A in the class, it is necessary for you to get an A on the final.
(d) You get an A on the final, but you don’t do every exercise in this book; nevertheless, you get an A in this class.
Class Problems
Problem 3.3.
When the mathematician says to his student, “If a function is not continuous, then it
is not differentiable,” then letting D stand for “differentiable” and C for continuous,
“mcs” — 2017/6/5 — 19:42 — page 70 — #78
70
Chapter 3
Logical Formulas
the only proper translation of the mathematician’s statement would be
NOT.C / IMPLIES NOT .D/;
or equivalently,
D IMPLIES C:
But when a mother says to her son, “If you don’t do your homework, then you
can’t watch TV,” then letting T stand for “can watch TV” and H for “do your
homework,” a reasonable translation of the mother’s statement would be
NOT .H / IFF NOT .T /;
or equivalently,
H IFF T:
Explain why it is reasonable to translate these two IF-THEN statements in different ways into propositional formulas.
Homework Problems
Problem 3.4.
Describe a simple procedure which, given a positive integer argument, n, produces
a width n array of truth-values whose rows would be all the possible truth-value
assignments for n propositional variables. For example, for n D 2, the array would
be:
T
T
F
F
T
F
T
F
Your description can be in English, or a simple program in some familiar language such as Python or Java. If you do write a program, be sure to include some
sample output.
Problem 3.5.
Sloppy Sam is trying to prove a certain proposition P . He defines two related
propositions Q and R, and then proceeds to prove three implications:
P IMPLIES Q;
He then reasons as follows:
Q IMPLIES R;
R IMPLIES P:
“mcs” — 2017/6/5 — 19:42 — page 71 — #79
3.7. References
71
If Q is true, then since I proved .Q IMPLIES R/, I can conclude that R
is true. Now, since I proved .R IMPLIES P /, I can conclude that P is
true. Similarly, if R is true, then P is true and so Q is true. Likewise,
if P is true, then so are Q and R. So any way you look at it, all three
of P; Q and R are true.
(a) Exhibit truth tables for
.P IMPLIES Q/ AND .Q IMPLIES R/ AND .R IMPLIES P /
(*)
P AND Q AND R:
(**)
and for
Use these tables to find a truth assignment for P; Q; R so that (*) is T and (**) is F.
(b) You show these truth tables to Sloppy Sam and he says “OK, I’m wrong that
P; Q and R all have to be true, but I still don’t see the mistake in my reasoning.
Can you help me understand my mistake?” How would you explain to Sammy
where the flaw lies in his reasoning?
Problems for Section 3.2
Class Problems
Problem 3.6.
Propositional logic comes up in digital circuit design using the convention that T
corresponds to 1 and F to 0. A simple example is a 2-bit half-adder circuit. This
circuit has 3 binary inputs, a1 ; a0 and b, and 3 binary outputs, c; s1 ; s0 . The 2-bit
word a1 a0 gives the binary representation of an integer k between 0 and 3. The
3-bit word cs1 s0 gives the binary representation of k C b. The third output bit c is
called the final carry bit.
So if k and b were both 1, then the value of a1 a0 would be 01 and the value of
the output cs1 s0 would 010, namely, the 3-bit binary representation of 1 C 1.
In fact, the final carry bit equals 1 only when all three binary inputs are 1, that is,
when k D 3 and b D 1. In that case, the value of cs1 s0 is 100, namely, the binary
representation of 3 C 1.
“mcs” — 2017/6/5 — 19:42 — page 72 — #80
72
Chapter 3
Logical Formulas
This 2-bit half-adder could be described by the following formulas:
c0 D b
s0 D a0 XOR c0
c1 D a0 AND c0
the carry into column 1
s1 D a1 XOR c1
c2 D a1 AND c1
the carry into column 2
c D c2 :
(a) Generalize the above construction of a 2-bit half-adder to an n C 1 bit halfadder with inputs an ; : : : ; a1 ; a0 and b and outputs c; sn ; : : : ; s1 ; s0 . That is, give
simple formulas for si and ci for 0 i n C 1, where ci is the carry into column
i C 1, and c D cnC1 .
(b) Write similar definitions for the digits and carries in the sum of two n C 1-bit
binary numbers an : : : a1 a0 and bn : : : b1 b0 .
Visualized as digital circuits, the above adders consist of a sequence of singledigit half-adders or adders strung together in series. These circuits mimic ordinary
pencil-and-paper addition, where a carry into a column is calculated directly from
the carry into the previous column, and the carries have to ripple across all the
columns before the carry into the final column is determined. Circuits with this
design are called ripple-carry adders. Ripple-carry adders are easy to understand
and remember and require a nearly minimal number of operations. But the higherorder output bits and the final carry take time proportional to n to reach their final
values.
(c) How many of each of the propositional operations does your adder from part (b)
use to calculate the sum?
Homework Problems
Problem 3.7.
As in Problem 3.6, a digital circuit is called an .n C 1/-bit half-adder when it has
with n C 2 inputs
an ; : : : ; a1 ; a0 ; b
and n C 2 outputs
c; sn ; : : : ; s1 ; s0 :
The input-output specification of the half-adder is that, if the 0-1 values of inputs
an ; : : : ; a1 ; a0 are taken to be the .n C 1/-bit binary representation of an integer k
“mcs” — 2017/6/5 — 19:42 — page 73 — #81
3.7. References
73
then the 0-1 values of the outputs c; sn ; : : : ; s1 ; s0 are supposed to be the .nC2/-bit
binary representation of k C b.
For example suppose n D 2 and the values of a2 a1 a0 were 101. This is the
binary representation of k D 5. Now if the value of b was 1, then the output should
be the 4-bit representation of 5 C 1 D 6. Namely, the values of cs2 s1 s0 would be
0110.
There are many different circuit designs for half adders. The most straighforward
one is the “ripple carry” design described in Problem 3.6. We will now develop
a different design for a half-adder circuit called a parallel-design or “look-ahead
carry” half-adder. This design works by computing the values of higher-order digits
for both a carry of 0 and a carry of 1, in parallel. Then, when the carry from the
low-order digits finally arrives, the pre-computed answer can be quickly selected.
We’ll illustrate this idea by working out a parallel design for an .n C 1/-bit halfadder.
Parallel-design half-adders are built out of parallel-design circuits called add1modules. The input-output behavior of an add1-module is just a special case of a
half-adder, where instead of an adding an input b to the input, the add1-module
always adds 1. That is, an .n C 1/-bit add1-module has .n C 1/ binary inputs
an ; : : : ; a1 ; a0 ;
and n C 2 binary outputs
c pn ; : : : ; p1 ; p0 :
If an : : : a1 a0 are taken to be the .n C 1/-bit representation of an integer k then
cpn : : : p1 p0 is supposed to be the .n C 2/-bit binary representation of k C 1.
So a 1-bit add1-module just has input a0 and outputs c; p0 where
p0 WWD a0 XOR 1; .or more simply, p0 WWD NOT.a0 //;
c WWD a0 :
In the ripple-carry design, a double-size half-adder with 2.n C 1/ inputs takes
twice as long to produce its output values as an .n C 1/-input ripple-carry circuit.
With parallel-design add1-modules, a double-size add1-module produces its output
values nearly as fast as a single-size add1-modules. To see how this works, suppose
the inputs of the double-size module are
a2nC1 ; : : : ; a1 ; a0
and the outputs are
c; p2nC1 ; : : : ; p1 ; p0 :
“mcs” — 2017/6/5 — 19:42 — page 74 — #82
74
Chapter 3
Logical Formulas
We will build the double-size add1-module by having two single-size add1-modules
work in parallel. The setup is illustrated in Figure 3.1.
Namely, the first single-size add1-module handles the first n C 1 inputs. The inputs to this module are the low-order n C 1 input bits an ; : : : ; a1 ; a0 , and its outputs
will serve as the first n C 1 outputs pn ; : : : ; p1 ; p0 of the double-size module. Let
c.1/ be the remaining carry output from this module.
The inputs to the second single-size module are the higher-order n C 1 input bits
a2nC1 ; : : : ; anC2 ; anC1 . Call its first n C 1 outputs rn ; : : : ; r1 ; r0 and let c.2/ be its
carry.
(a) Write a formula for the carry c of the double-size add1-module solely in terms
of carries c.1/ and c.2/ of the single-size add1-modules.
(b) Complete the specification of the double-size add1-module by writing propositional formulas for the remaining outputs pnCi for 1 i n C 1. The formula
for pnCi should only involve the variables anCi , ri 1 and c.1/ .
(c) Explain how to build an .nC1/-bit parallel-design half-adder from an .nC1/bit add1-module by writing a propositional formula for the half-adder output si
using only the variables ai , pi and b.
(d) The speed or latency of a circuit is determined by the largest number of gates
on any path from an input to an output. In an n-bit ripple carry circuit(Problem 3.6),
there is a path from an input to the final carry output that goes through about 2n
gates. In contrast, parallel half-adders are exponentially faster than ripple-carry
half-adders. Confirm this by determining the largest number of propositional operations, that is, gates, on any path from an input to an output of an n-bit add1-module.
(You may assume n is a power of 2.)
Exam Problems
Problem 3.8.
Claim. There are exactly two truth environments (assignments) for the variables
M; N; P; Q; R; S that satisfy the following formula:
.P OR Q/ AND .Q OR R/ AND .R OR S / AND .S OR P / AND M AND N
„ ƒ‚ …
„ ƒ‚ …
„ ƒ‚ …
„ ƒ‚ …
clause (1)
clause (2)
clause (3)
clause (4)
(a) This claim could be proved by truth-table. How many rows would the truth
table have?
“mcs” — 2017/6/5 — 19:42 — page 75 — #83
3.7. References
75
a2nC1
.nC1/-bit add1
c.2/
rn
c
anC2 anC1
r1
a1
an
a0
.nC1/-bit add1
c.1/
r0
2.nC2/-bit add1 module
p2nC1
p nC2 pnC1
pn
Figure 3.1 Structure of a Double-size add1 Module.
p1
p0
“mcs” — 2017/6/5 — 19:42 — page 76 — #84
76
Chapter 3
Logical Formulas
(b) Instead of a truth-table, prove this claim with an argument by cases according
to the truth value of P .
Problem 3.9.
An n-bit AND-circuit has 0-1 valued inputs a0 ; a1 ; : : : ; an
whose value will be
c D a0 AND a1 AND AND an
1
and one output c
1:
There are various ways to design an n-bit AND-circuit. A serial design is simply
a series of AND-gates, each with one input being a circuit input ai and the other
input being the output of the previous gate as shown in Figure 3.2.
We can also use a tree design. A 1-bit tree design is just a wire, that is c WWD a1 .
Assuming for simplicity that n is a power of two, an n-input tree circuit for n > 1
simply consists of two n=2-input tree circuits whose outputs are AND’d to produce
output c, as in Figure 3.3. For example, a 4-bit tree design circut is shown in
Figure 3.4.
(a) How many AND-gates are in the n-input serial circuit?
(b) The “speed” or latency of a circuit is the largest number of gates on any path
from an input to an output. Briefly explain why the tree circuit is exponentially
faster than the serial circuit.
(c) Assume n is a power of two. Prove that the n-input tree circuit has n
AND -gates.
1
Problems for Section 3.3
Practice Problems
Problem 3.10.
Indicate whether each of the following propositional formulas is valid (V), satisfiable but not valid (S), or not satisfiable (N). For the satisfiable ones, indicate a
satisfying truth assignment.
“mcs” — 2017/6/5 — 19:42 — page 77 — #85
3.7. References
77
Figure 3.2 A serial AND-circuit.
“mcs” — 2017/6/5 — 19:42 — page 78 — #86
78
Chapter 3
Logical Formulas
Figure 3.3
An n-bit AND-tree circuit.
Figure 3.4 A 4-bit AND-tree circuit.
“mcs” — 2017/6/5 — 19:42 — page 79 — #87
3.7. References
79
M IMPLIES Q
M IMPLIES .P OR Q/
M IMPLIES ŒM AND .P IMPLIES M /
.P OR Q/ IMPLIES Q
.P OR Q/ IMPLIES .P AND Q/
.P OR Q/ IMPLIES ŒM AND .P IMPLIES M /
.P XOR Q/ IMPLIES Q
.P XOR Q/ IMPLIES .P OR Q/
.P XOR Q/ IMPLIES ŒM AND .P IMPLIES M /
Problem 3.11.
Show truth tables that verify the equivalence of the following two propositional
formulas
.P XOR Q/;
NOT .P IFF
Q/:
Problem 3.12.
Prove that the propositional formulas
P OR Q OR R
and
.P AND NOT.Q// OR .Q AND NOT.R// OR .R AND NOT.P // OR .P AND Q AND R/:
are equivalent.
Problem 3.13.
Prove by truth table that OR distributes over AND, namely,
P OR .Q AND R/
is equivalent to .P OR Q/ AND .P OR R/
(3.31)
“mcs” — 2017/6/5 — 19:42 — page 80 — #88
80
Chapter 3
Logical Formulas
Exam Problems
Problem 3.14.
The formula
NOT.A IMPLIES
B/ AND A AND C
IMPLIES
D AND E AND F AND G AND H AND I AND J AND K AND L AND M
turns out to be valid.
(a) Explain why verifying the validity of this formula by truth table would be very
hard for one person to do with pencil and paper (no computers).
(b) Verify that the formula is valid, reasoning by cases according to the truth value
of A.
Proof. Case: (A is True).
Case: (A is False).
Class Problems
Problem 3.15. (a) Verify by truth table that
.P IMPLIES Q/ OR .Q IMPLIES P /
is valid.
(b) Let P and Q be propositional formulas. Describe a single formula R using
only AND’s, OR’s, NOT’s, and copies of P and Q, such that R is valid iff P and Q
are equivalent.
(c) A propositional formula is satisfiable iff there is an assignment of truth values
to its variables—an environment—that makes it true. Explain why
P is valid
iff
NOT .P /
is not satisfiable.
(d) A set of propositional formulas P1 ; : : : ; Pk is consistent iff there is an environment in which they are all true. Write a formula S such that the set P1 ; : : : ; Pk
is not consistent iff S is valid.
“mcs” — 2017/6/5 — 19:42 — page 81 — #89
3.7. References
81
Problem 3.16.
This problem4 examines whether the following specifications are satisfiable:
1. If the file system is not locked, then
(a) new messages will be queued.
(b) new messages will be sent to the messages buffer.
(c) the system is functioning normally, and conversely, if the system is
functioning normally, then the file system is not locked.
2. If new messages are not queued, then they will be sent to the messages buffer.
3. New messages will not be sent to the message buffer.
(a) Begin by translating the five specifications into propositional formulas using
four propositional variables:
L WWD file system locked;
Q WWD new messages are queued;
B WWD new messages are sent to the message buffer;
N WWD system functioning normally:
(b) Demonstrate that this set of specifications is satisfiable by describing a single
truth assignment for the variables L; Q; B; N and verifying that under this assignment, all the specifications are true.
(c) Argue that the assignment determined in part (b) is the only one that does the
job.
Problems for Section 3.4
Practice Problems
Problem 3.17.
A half dozen different operators may appear in propositional formulas, but just
AND , OR , and NOT are enough to do the job. That is because each of the operators
is equivalent to a simple formula using only these three operators. For example,
A IMPLIES B is equivalent to NOT.A/ OR B. So all occurences of IMPLIES in a
formula can be replaced using just NOT and OR.
4 Revised
from Rosen, 5th edition, Exercise 1.1.36
“mcs” — 2017/6/5 — 19:42 — page 82 — #90
82
Chapter 3
Logical Formulas
(a) Write formulas using only AND, OR, NOT that are equivalent to each of A IFF B
and A XOR B. Conclude that every propositional formula is equivalent to an ANDOR - NOT formula.
(b) Explain why you don’t even need AND.
(c) Explain how to get by with the single operator NAND where A NAND B is
equivalent by definition to NOT.A AND B/.
Class Problems
Problem 3.18.
The propositional connective NOR is defined by the rule
P NOR Q WWD .NOT.P / AND NOT.Q//:
Explain why every propositional formula—possibly involving any of the usual operators such as IMPLIES, XOR, . . . —is equivalent to one whose only connective is
NOR .
Problem 3.19.
Explain how to read off a conjunctive form for a propositional formula directly
from a disjunctive form for its complement.
Problem 3.20.
Let P be the proposition depending on propositional variable A; B; C; D whose
truth values for each truth assignment to A; B; C; D are given in the table below.
Write out both a disjunctive and a conjunctive normal form for P .
“mcs” — 2017/6/5 — 19:42 — page 83 — #91
3.7. References
83
A
T
T
T
T
T
T
T
T
F
F
F
F
F
F
F
F
B C D P
T T T T
T T F F
T F T T
T F F F
F T T T
F T F T
F F T T
F F F T
T T T T
T T F F
T F T T
T F F F
F T T F
F T F F
F F T T
F F F T
Homework Problems
Problem 3.21.
Use the equivalence axioms of Section 3.4.2 to convert the formula
A XOR B XOR C
(a) . . . to disjunctive (OR of AND’s) form,
(b) . . . to conjunctive (AND of OR’s) form.
Problems for Section 3.5
Class Problems
Problem 3.22.
The circuit-SAT problem is the problem of determining, for any given digital circuit
with one output wire, whether there are truth values that can be fed into the circuit
input wires which will lead the circuit to give output T.
“mcs” — 2017/6/5 — 19:42 — page 84 — #92
84
Chapter 3
Logical Formulas
It’s easy to see that any efficient way of solving the circuit-SAT problem would
yield an efficient way to solve the usual SAT problem for propositional formulas
(Section 3.5). Namely, for any formula F , just construct a circuit CF using that
computes the values of the formula. Then there are inputs for which CF gives
output true iff F is satisfiable. Constructing CF from F is easy, using a binary
gate in CF for each propositional connective in F . So an efficient circuit-SAT
procedure leads to an efficient SAT procedure.
Conversely, there is a simple recursive procedure that will construct, given C , a
formula EC that is equivalent to C in the sense that the truth value EC and the output of C are the same for every truth assignment of the variables. The difficulty is
that, in general, the “equivalent” formula EC , will be exponentially larger than C .
For the purposes of showing that satifiability of circuits and satisfiability of formulas take roughly the same effort to solve, spending an exponential time translating
one problem to the other swamps any benefit in switching from one problem to the
other.
So instead of a formula EC that is equivalent to C , we aim instead for a formula
FC that is “equisatisfiable” with C . That is, there will be input values that make
C output True iff there is a truth assignment that satisfies FC . (In fact, FC and C
need not even use the same variables.) But now we make sure that the amount of
computation needed to construct FC is not much larger than the size of the circuit
C . In particular, the size of FC will also not be much larger than C .
The idea behind the construction of FC is that, given any digital circuit C with
binary gates and one output, we can assign a distinct variable to each wire of C .
Then for each gate of C , we can set up a propositional formula that represents the
constraints that the gate places on the values of its input and output wires. For
example, for an AND gate with input wire variables P and Q and output wire
variable R, the constraint proposition would be
.P AND Q/ IFF R:
(3.32)
(a) Given a circuit C , explain how to easily find a formula FC of size proportional
to the number of wires in C such that FC is satisfiable iff C gives output T for some
set of input values.
(b) Conclude that any efficient way of solving SAT would yield an efficient way
to solve circuit-SAT.
Homework Problems
Problem 3.23.
A 3-conjunctive form (3CF) formula is a conjunctive form formula in which each
“mcs” — 2017/6/5 — 19:42 — page 85 — #93
3.7. References
85
OR -term
is an OR of at most 3 variables or negations of variables. Although it
may be hard to tell if a propositional formula F is satisfiable, it is always easy to
construct a formula C.F / that is
in 3-conjunctive form,
has at most 24 times as many occurrences of variables as F , and
is satisfiable iff F is satisfiable.
To construct C.F /, introduce a different new variables for each operator that
occurs in F . For example, if F was
..P XOR Q/ XOR R/ OR .P AND S /
(3.33)
we might use new variables X1 , X2 O and A corresponding to the operator occurrences as follows:
..P „ƒ‚…
XOR Q/ XOR R/ OR .P „ƒ‚…
AND S /:
„ƒ‚… „ƒ‚…
X1
X2
O
A
Next we write a formula that constrains each new variable to have the same truth
value as the subformula determined by its corresponding operator. For the example
above, these constraining formulas would be
X1 IFF .P XOR Q/;
X2 IFF .X1 XOR R/;
A IFF .P AND S /;
O IFF .X2 OR A/
(a) Explain why the AND of the four constraining formulas above along with a
fifth formula consisting of just the variable O will be satisfiable iff (3.33) is satisfiable.
(b) Explain why each constraining formula will be equivalent to a 3CF formula
with at most 24 occurrences of variables.
(c) Using the ideas illustrated in the previous parts, explain how to construct C.F /
for an arbitrary propositional formula F .
“mcs” — 2017/6/5 — 19:42 — page 86 — #94
86
Chapter 3
Logical Formulas
Problems for Section 3.6
Practice Problems
Problem 3.24.
For each of the following propositions:
1. 8x 9y: 2x
yD0
2. 8x 9y: x
2y D 0
3. 8x: x < 10 IMPLIES .8y: y < x IMPLIES y < 9/
4. 8x 9y: Œy > x ^ 9z: y C z D 100
determine which propositions are true when the variables range over:
(a) the nonnegative integers.
(b) the integers.
(c) the real numbers.
Problem 3.25.
Let Q.x; y/ be the statement
“x has been a contestant on television show y.”
The universe of discourse for x is the set of all students at your school and for y is
the set of all quiz shows that have ever been on television.
Determine whether or not each of the following expressions is logically equivalent to the sentence:
“No student at your school has ever been a contestant on a television quiz show.”
(a) 8x 8y: NOT.Q.x; y//
(b) 9x 9y: NOT.Q.x; y//
(c) NOT.8x 8y: Q.x; y//
(d) NOT.9x 9y: Q.x; y//
“mcs” — 2017/6/5 — 19:42 — page 87 — #95
3.7. References
87
Problem 3.26.
Express each of the following statements using quantifiers, logical connectives,
and/or the following predicates
P .x/: x is a monkey,
Q.x/: x is a 6.042 TA,
R.x/: x comes from the 23rd century,
S.x/: x likes to eat pizza,
where x ranges over all living things.
(a) No monkeys like to eat pizza.
(b) Nobody from the 23rd century dislikes eating pizza.
(c) All 6.042 TAs are monkeys.
(d) No 6.042 TA comes from the 23rd century.
(e) Does part (d) follow logically from parts (a), (b), (c)? If so, give a proof. If
not, give a counterexample.
(f) Translate into English: .8x/.R.x/ _ S.x/ ! Q.x//.
(g) Translate into English:
Œ9x: R.x/ AND NOT.Q.x// IMPLIES 8x: .P .x/ IMPLIES S.x//:
Problem 3.27.
Find a counter-model showing the following is not valid.
9x:P .x/ IMPLIES 8x:P .x/
(Just define your counter-model. You do not need to verify that it is correct.)
Problem 3.28.
Find a counter-model showing the following is not valid.
Œ9x: P .x/ AND 9x:Q.x/ IMPLIES 9x:ŒP .x/ AND Q.x/
(Just define your counter-model. You do not need to verify that it is correct.)
“mcs” — 2017/6/5 — 19:42 — page 88 — #96
88
Chapter 3
Logical Formulas
Problem 3.29.
Which of the following are valid? For those that are not valid, desribe a countermodel.
(a) 9x9y: P .x; y/ IMPLIES 9y9x: P .x; y/
(b) 8x9y: Q.x; y/ IMPLIES 9y8x: Q.x; y/
(c) 9x8y: R.x; y/ IMPLIES 8y9x: R.x; y/
(d) NOT.9x S.x// IFF 8x NOT.S.x//
Problem 3.30. (a) Verify that the propositional formula
.P IMPLIES Q/ OR .Q IMPLIES P /
is valid.
(b) The valid formula of part (a) leads to sound proof method: to prove that an implication is true, just prove that its converse is false.5 For example, from elementary
calculus we know that the assertion
If a function is continuous, then it is differentiable
is false. This allows us to reach at the correct conclusion that its converse,
If a function is differentiable, then it is continuous
is true, as indeed it is.
But wait a minute! The implication
If a function is differentiable, then it is not continuous
is completely false. So we could conclude that its converse
If a function is not continuous, then it is differentiable,
should be true, but in fact the converse is also completely false.
So something has gone wrong here. Explain what.
5 This
problem was stimulated by the discussion of the fallacy in [4].
“mcs” — 2017/6/5 — 19:42 — page 89 — #97
3.7. References
89
Class Problems
Problem 3.31.
A media tycoon has an idea for an all-news television network called LNN: The
Logic News Network. Each segment will begin with a definition of the domain of
discourse and a few predicates. The day’s happenings can then be communicated
concisely in logic notation. For example, a broadcast might begin as follows:
THIS IS LNN. The domain of discourse is
fAlbert; Ben; Claire; David; Emilyg:
Let D.x/ be a predicate that is true if x is deceitful. Let L.x; y/
be a predicate that is true if x likes y. Let G.x; y/ be a predicate that
is true if x gave gifts to y.
Translate the following broadcasts in logic notation into (English) statements.
(a)
NOT.D.Ben/ OR
D.David// IMPLIES
.L.Albert; Ben/ AND L.Ben; Albert//:
(b)
8x: ..x D Claire AND NOT.L.x; Emily/// OR .x ¤ Claire AND L.x; Emily///
AND
8x: ..x D David AND L.x; Claire// OR .x ¤ David AND NOT.L.x; Claire////
(c)
NOT .D.Claire// IMPLIES
.G.Albert; Ben/ AND 9x: G.Ben; x//
(d)
8x9y9z .y ¤ z/ AND L.x; y/ AND NOT.L.x; z//:
(e) How could you express “Everyone except for Claire likes Emily” using just
propositional connectives without using any quantifiers (8; 9)? Can you generalize
to explain how any logical formula over this domain of discourse can be expressed
without quantifiers? How big would the formula in the previous part be if it was
expressed this way?
“mcs” — 2017/6/5 — 19:42 — page 90 — #98
90
Chapter 3
Logical Formulas
Problem 3.32.
The goal of this problem is to translate some assertions about binary strings into
logic notation. The domain of discourse is the set of all finite-length binary strings:
, 0, 1, 00, 01, 10, 11, 000, 001, . . . . (Here denotes the empty string.) In your
translations, you may use all the ordinary logic symbols (including =), variables,
and the binary symbols 0, 1 denoting 0, 1.
A string like 01x0y of binary symbols and variables denotes the concatenation
of the symbols and the binary strings represented by the variables. For example, if
the value of x is 011 and the value of y is 1111, then the value of 01x0y is the
binary string 0101101111.
Here are some examples of formulas and their English translations. Names for
these predicates are listed in the third column so that you can reuse them in your
solutions (as we do in the definition of the predicate NO -1 S below).
Meaning
x is a prefix of y
x is a substring of y
x is empty or a string of 0’s
Formula
9z .xz D y/
9u9v .uxv D y/
NOT . SUBSTRING .1; x//
Name
PREFIX (x; y)
SUBSTRING (x; y)
NO -1 S (x)
(a) x consists of three copies of some string.
(b) x is an even-length string of 0’s.
(c) x does not contain both a 0 and a 1.
(d) x is the binary representation of 2k C 1 for some integer k 0.
(e) An elegant, slightly trickier way to define NO -1 S.x/ is:
PREFIX .x; 0x/:
(*)
Explain why (*) is true only when x is a string of 0’s.
Problem 3.33.
For each of the logical formulas, indicate whether or not it is true when the domain of discourse is N, (the nonnegative integers 0, 1, 2, . . . ), Z (the integers), Q
(the rationals), R (the real numbers), and C (the complex numbers). Add a brief
explanation to the few cases that merit one.
“mcs” — 2017/6/5 — 19:42 — page 91 — #99
3.7. References
91
9x: x 2 D 2
8x:9y: x 2 D y
8y:9x: x 2 D y
8x ¤ 0:9y: xy D 1
9x:9y: x C 2y D 2 AND 2x C 4y D 5
Problem 3.34.
Show that
.8x9y: P .x; y// ! 8z: P .z; z/
is not valid by describing a counter-model.
Homework Problems
Problem 3.35.
Express each of the following predicates and propositions in formal logic notation.
The domain of discourse is the nonnegative integers, N. Moreover, in addition to
the propositional operators, variables and quantifiers, you may define predicates
using addition, multiplication, and equality symbols, and nonnegative integer constants (0, 1,. . . ), but no exponentiation (like x y ). For example, the predicate “n is
an even number” could be defined by either of the following formulas:
9m: .2m D n/;
9m: .m C m D n/:
(a) m is a divisor of n.
(b) n is a prime number.
(c) n is a power of a prime.
Problem 3.36.
Translate the following sentence into a predicate formula:
There is a student who has e-mailed at most two other people in the
class, besides possibly himself.
The domain of discourse should be the set of students in the class; in addition,
the only predicates that you may use are
“mcs” — 2017/6/5 — 19:42 — page 92 — #100
92
Chapter 3
Logical Formulas
equality, and
E.x; y/, meaning that “x has sent e-mail to y.”
Problem 3.37. (a) Translate the following sentence into a predicate formula:
There is a student who has e-mailed at most n other people in the class,
besides possibly himself.
The domain of discourse should be the set of students in the class; in addition, the
only predicates that you may use are
equality,
E.x; y/, meaning that “x has sent e-mail to y.”
(b) Explain how you would use your predicate formula (or some variant of it) to
express the following two sentences.
1. There is a student who has emailed at least n other people in the class, besides
possibly himself.
2. There is a student who has emailed exactly n other people in the class, besides
possibly himself.
Exam Problems
Problem 3.38.
For each of the logic formulas below, indicate the smallest domain in which it is
true, among
N(nonnegative integers); Z(integers); Q(rationals); R(reals); C(complex numbers);
or state “none” if it is not true in any of them.
i. 8x9y: y D 3x
ii. 8x9y: 3y D x
iii. 8x9y: y 2 D x
iv. 8x9y: y < x
v. 8x9y: y 3 D x
“mcs” — 2017/6/5 — 19:42 — page 93 — #101
3.7. References
93
vi. 8x ¤ 0: 9y; z: y ¤ z AND y 2 D x D z 2
Problem 3.39.
The following predicate logic formula is invalid:
8x; 9y:P .x; y/ ! 9y; 8x:P .x; y/
Which of the following are counter models for it?
1. The predicate P .x; y/ D ‘y x D 1’ where the domain of discourse is Q.
2. The predicate P .x; y/ D ‘y < x’ where the domain of discourse is R.
3. The predicate P .x; y/ D ‘y x D 2’ where the domain of discourse is R
without 0.
4. The predicate P .x; y/ D ‘yxy D x’ where the domain of discourse is the
set of all binary strings, including the empty string.
Problem 3.40.
Some students from a large class will be lined up left to right. There will be at least
two students in the line. Translate each of the following assertions into predicate
formulas with the set of students in the class as the domain of discourse. The only
predicates you may use are
equality and,
F .x; y/, meaning that “x is somewhere to the left of y in the line.” For
example, in the line “CDA”, both F .C; A/ and F .C; D/ are true.
Once you have defined a formula for a predicate P you may use the abbreviation
“P ” in further formulas.
(a) Student x is in the line.
(b) Student x is first in line.
(c) Student x is immediately to the right of student y.
(d) Student x is second.
“mcs” — 2017/6/5 — 19:42 — page 94 — #102
94
Chapter 3
Logical Formulas
Problem 3.41.
We want to find predicate formulas about the nonnegative integers N in which is
the only predicate that appears, and no constants appear.
For example, there is such a formula defining the equality predicate:
Œx D y WWD Œx y AND y x:
Once predicate is shown to be expressible solely in terms of , it may then be used
in subsequent translations. For example,
Œx > 0 WWD 9y: NOT.x D y/ AND y x:
(a) Œx D 0.
(b) Œx D y C 1.
Hint: If an integer is bigger than y, then it must be x.
(c) x D 3.
Problem 3.42.
Predicate Formulas whose only predicate symbol is equality are called “pure equality” formulas. For example,
8x 8y: x D y
(1-element)
is a pure equality formula. Its meaning is that there is exactly one element in the
domain of discourse.6 Another such formula is
9a 9b 8x: x D a OR x D b:
( 2-elements)
Its meaning is that there are at most two elements in the domain of discourse.
A formula that is not a pure equality formula is
x y:
(not-pure)
Formula (not-pure) uses the less-than-or-equal predicate which is not allowed.7
(a) Describe a pure equality formula that means that there are exactly two elements in the domain of discourse.
6 Remember,
a domain of discourse is not allowed to be empty.
fact, formula (not-pure) only makes sense when the domain elements are ordered, while pure
equality formulas make sense over every domain.
7 In
“mcs” — 2017/6/5 — 19:42 — page 95 — #103
3.7. References
95
(b) Describe a pure equality formula that means that there are exactly three elements in the domain of discourse.
Problem 3.43. (a) A predicate R on the nonnegative integers is true infinitely often
(i.o.) when R.n/ is true for infinitely many n 2 N.
We can express the fact that R is true i.o. with a formula of the form:
Q1 Q2 : R.n/;
where Q1 ; Q2 are quantifiers from among
8n; 9n; 8n n0 ; 9n n0 ;
8n0 ; 9n0 ; 8n0 n; 9n0 n;
and n; n0 range over nonnegative integers.
Identify the proper quantifers:
Q1
Q2
(b) A predicate S on the nonnegative integers is true almost everywhere (a.e.)
when S.n/ is false for only finitely many n 2 N.
We can express the fact that S is true a.e. with a formula of the form
Q3 Q4 : S.n/;
where Q3 ; Q4 are quantifiers from those above:
8n; 9n; 8n n0 ; 9n n0 ;
8n0 ; 9n0 ; 8n0 n; 9n0 n:
Identify the proper quantifers:
Q3
Q4
Problem 3.44.
Let f W N ! R be a real-valued total function. A limit point of f is a real number
“mcs” — 2017/6/5 — 19:42 — page 96 — #104
96
Chapter 3
Logical Formulas
r 2 R such that f .n/ is close to r for infinitely many n, where “close to” means
within distance for whatever positive real number you may choose.
We can express the fact that r is a limit point of f with a logical formula of the
form:
Q0 Q1 Q2 : jf .n/ rj ;
where Q0 ; Q1 ; Q2 is a sequence of three quantifiers from among:
8n;
9n;
8n n0 ; 9n n0 :
8n0 ;
9n0 ;
8n0 n; 9n0 n:
8 0; 9 0; 8 > 0;
9 > 0:
Here the n; n0 range over nonnegative integers, and ranges over real numbers.
Identify the proper quantifers:
Q0
Q1
Q2
“mcs” — 2017/6/5 — 19:42 — page 97 — #105
4
Mathematical Data Types
We have assumed that you’ve already been introduced to the concepts of sets, sequences, and functions, and we’ve used them informally several times in previous
sections. In this chapter, we’ll now take a more careful look at these mathematical data types. We’ll quickly review the basic definitions, add a few more such as
“images” and “inverse images” that may not be familiar, and end the chapter with
some methods for comparing the sizes of sets.
4.1
Sets
Informally, a set is a bunch of objects, which are called the elements of the set.
The elements of a set can be just about anything: numbers, points in space, or even
other sets. The conventional way to write down a set is to list the elements inside
curly-braces. For example, here are some sets:
A D fAlex; Tippy; Shells; Shadowg
B D fred; blue; yellowg
C D ffa; bg; fa; cg; fb; cgg
dead pets
primary colors
a set of sets
This works fine for small finite sets. Other sets might be defined by indicating how
to generate a list of them:
D WWD f1; 2; 4; 8; 16; : : :g
the powers of 2
The order of elements is not significant, so fx; yg and fy; xg are the same set
written two different ways. Also, any object is, or is not, an element of a given set—
there is no notion of an element appearing more than once in a set.1 So, writing
fx; xg is just indicating the same thing twice: that x is in the set. In particular,
fx; xg D fxg.
The expression “e 2 S” asserts that e is an element of set S. For example,
32 2 D and blue 2 B, but Tailspin 62 A—yet.
Sets are simple, flexible, and everywhere. You’ll find some set mentioned in
nearly every section of this text.
1 It’s
not hard to develop a notion of multisets in which elements can occur more than once, but
multisets are not ordinary sets and are not covered in this text.
“mcs” — 2017/6/5 — 19:42 — page 98 — #106
98
Chapter 4
4.1.1
Mathematical Data Types
Some Popular Sets
Mathematicians have devised special symbols to represent some common sets.
symbol
;
N
Z
Q
R
C
set
the empty set
nonnegative integers
integers
rational numbers
real numbers
complex numbers
elements
none
f0; 1; 2; 3; : : :g
f: : : ; 3; 2; 1; 0; 1; 2; 3; : : :g
1
5
2;
3 ; 16;petc.
; e; p9; 2; etc.
i; 19
2 2i; etc.
2 ;
A superscript “C ” restricts a set to its positive elements; for example, RC denotes
the set of positive real numbers. Similarly, Z denotes the set of negative integers.
4.1.2
Comparing and Combining Sets
The expression S T indicates that set S is a subset of set T , which means that
every element of S is also an element of T . For example, N Z because every
nonnegative integer is an integer; Q R because every rational number is a real
number, but C 6 R because not every complex number is a real number.
As a memory trick, think of the “” symbol as like the “” sign with the smaller
set or number on the left-hand side. Notice that just as n n for any number n,
also S S for any set S .
There is also a relation on sets like the “less than” relation < on numbers.
S T means that S is a subset of T , but the two are not equal. So just as n 6< n
for every number n, also A 6 A, for every set A. “S T ” is read as “S is a strict
subset of T .”
There are several basic ways to combine sets. For example, suppose
X WWD f1; 2; 3g;
Y WWD f2; 3; 4g:
Definition 4.1.1.
The union of sets A and B, denoted A [ B, includes exactly the elements
appearing in A or B or both. That is,
x 2A[B
So X [ Y D f1; 2; 3; 4g.
IFF
x 2 A OR x 2 B:
“mcs” — 2017/6/5 — 19:42 — page 99 — #107
4.1. Sets
99
The intersection of A and B, denoted A \ B, consists of all elements that
appear in both A and B. That is,
x 2A\B
x 2 A AND x 2 B:
IFF
So, X \ Y D f2; 3g.
The set difference of A and B, denoted A
are in A, but not in B. That is,
x2A
So, X
Y D f1g and Y
B
B, consists of all elements that
x 2 A AND x … B:
IFF
X D f4g.
Often all the sets being considered are subsets of a known domain of discourse
D. Then for any subset A of D, we define A to be the set of all elements of D not
in A. That is,
A WWD D A:
The set A is called the complement of A. So
A D ; IFF A D D:
For example, if the domain we’re working with is the integers, the complement
of the nonnegative integers is the set of negative integers:
NDZ :
We can use complement to rephrase subset in terms of equality
A B is equivalent to A \ B D ;:
4.1.3
Power Set
The set of all the subsets of a set A is called the power set pow.A/ of A. So
B 2 pow.A/
IFF
B A:
For example, the elements of pow.f1; 2g/ are ;; f1g; f2g and f1; 2g.
More generally, if A has n elements, then there are 2n sets in pow.A/—see Theorem 4.5.5. For this reason, some authors use the notation 2A instead of pow.A/.
“mcs” — 2017/6/5 — 19:42 — page 100 — #108
100
Chapter 4
4.1.4
Mathematical Data Types
Set Builder Notation
An important use of predicates is in set builder notation. We’ll often want to talk
about sets that cannot be described very well by listing the elements explicitly or
by taking unions, intersections, etc., of easily described sets. Set builder notation
often comes to the rescue. The idea is to define a set using a predicate; in particular,
the set consists of all values that make the predicate true. Here are some examples
of set builder notation:
A WWD fn 2 N j n is a prime and n D 4k C 1 for some integer kg;
B WWD fx 2 R j x 3
3x C 1 > 0g;
C WWD fa C bi 2 C j a2 C 2b 2 1g;
D WWD fL 2 books j L is cited in this textg:
The set A consists of all nonnegative integers n for which the predicate
“n is a prime and n D 4k C 1 for some integer k”
is true. Thus, the smallest elements of A are:
5; 13; 17; 29; 37; 41; 53; 61; 73; : : : :
Trying to indicate the set A by listing these first few elements wouldn’t work very
well; even after ten terms, the pattern is not obvious. Similarly, the set B consists
of all real numbers x for which the predicate
x3
3x C 1 > 0
is true. In this case, an explicit description of the set B in terms of intervals would
require solving a cubic equation. Set C consists of all complex numbers a C bi
such that:
a2 C 2b 2 1
This is an oval-shaped region around the origin in the complex plane. Finally, the
members of set D can be determined by filtering out journal articles in from the list
of references in the Bibliography 22.5.
4.1.5
Proving Set Equalities
Two sets are defined to be equal if they have exactly the same elements. That is,
X D Y means that z 2 X if and only if z 2 Y , for all elements z.2 So, set
equalities can be formulated and proved as “iff” theorems. For example:
2 This
is actually the first of the ZFC axioms for set theory mentioned at the end of Section 1.3
and discussed further in Section 8.3.2.
“mcs” — 2017/6/5 — 19:42 — page 101 — #109
4.1. Sets
101
Theorem 4.1.2. [Distributive Law for Sets] Let A, B and C be sets. Then:
A \ .B [ C / D .A \ B/ [ .A \ C /
(4.1)
Proof. The equality (4.1) is equivalent to the assertion that
z 2 A \ .B [ C / iff
z 2 .A \ B/ [ .A \ C /
(4.2)
for all z. Now we’ll prove (4.2) by a chain of iff’s.
Now we have
z 2 A \ .B [ C /
iff
.z 2 A/ AND .z 2 B [ C /
(def of \)
iff
.z 2 A/ AND .z 2 B OR z 2 C /
(def of [)
iff
.z 2 A AND z 2 B/ OR .z 2 A AND z 2 C / (AND distributivity (3.9))
iff
.z 2 A \ B/ OR .z 2 A \ C /
(def of \)
iff
z 2 .A \ B/ [ .A \ C /
(def of [)
The proof of Theorem 4.1.2 illustrates a general method for proving a set equality
involving the basic set operations by checking that a corresponding propositional
formula is valid. As a further example, from De Morgan’s Law (3.14) for propositions
NOT .P AND Q/ is equivalent to P OR Q
we can derive (Problem 4.5) a corresponding De Morgan’s Law for set equality:
A \ B D A [ B:
(4.3)
Despite this correspondence between two kinds of operations, it’s important not
to confuse propositional operations with set operations. For example, if X and Y
are sets, then it is wrong to write “X AND Y ” instead of “X \ Y .” Applying AND
to sets will cause your compiler—or your grader—to throw a type error, because
an operation that is only supposed to be applied to truth values has been applied to
sets. Likewise, if P and Q are propositions, then it is a type error to write “P [ Q”
instead of “P OR Q.”
“mcs” — 2017/6/5 — 19:42 — page 102 — #110
102
4.2
Chapter 4
Mathematical Data Types
Sequences
Sets provide one way to group a collection of objects. Another way is in a sequence,
which is a list of objects called its components, members, or elements. Short sequences are commonly described by listing the elements between parentheses; for
example, the sequence .a; b; c/ has three components. It would also be referred to
as a three element sequence or a sequence of length three. These phrases are all
synonyms—sequences are so basic that they appear everywhere and there are a lot
of ways to talk about them.
While both sets and sequences perform a gathering role, there are several differences.
The elements of a set are required to be distinct, but elements in a sequence
can be the same. Thus, .a; b; a/ is a valid sequence of length three, but
fa; b; ag is a set with two elements, not three.
The elements in a sequence have a specified order, but the elements of a
set do not. For example, .a; b; c/ and .a; c; b/ are different sequences, but
fa; b; cg and fa; c; bg are the same set.
Texts differ on notation for the empty sequence; we use for the empty
sequence.
The product operation is one link between sets and sequences. A Cartesian
product of sets, S1 S2 Sn , is a new set consisting of all sequences where
the first component is drawn from S1 , the second from S2 , and so forth. Length two
sequences are called pairs.3 For example, N fa; bg is the set of all pairs whose
first element is a nonnegative integer and whose second element is an a or a b:
N fa; bg D f.0; a/; .0; b/; .1; a/; .1; b/; .2; a/; .2; b/; : : :g
A product of n copies of a set S is denoted S n . For example, f0; 1g3 is the set of
all 3-bit sequences:
f0; 1g3 D f.0; 0; 0/; .0; 0; 1/; .0; 1; 0/; .0; 1; 1/; .1; 0; 0/; .1; 0; 1/; .1; 1; 0/; .1; 1; 1/g
3 Some
texts call them ordered pairs.
“mcs” — 2017/6/5 — 19:42 — page 103 — #111
4.3. Functions
4.3
103
Functions
4.3.1
Domains and Images
A function assigns an element of one set, called the domain, to an element of another set, called the codomain. The notation
f WA!B
indicates that f is a function with domain A and codomain B. The familiar notation
“f .a/ D b” indicates that f assigns the element b 2 B to a. Here b would be
called the value of f at argument a.
Functions are often defined by formulas, as in:
f1 .x/ WWD
1
x2
where x is a real-valued variable, or
f2 .y; z/ WWD y10yz
where y and z range over binary strings, or
f3 .x; n/ WWD the length n sequence .x; : : : ; x/
„ ƒ‚ …
n x’s
where n ranges over the nonnegative integers.
A function with a finite domain could be specified by a table that shows the value
of the function at each element of the domain. For example, a function f4 .P; Q/
where P and Q are propositional variables is specified by:
P Q f4 .P; Q/
T T
T
T F
F
F T
T
F F
T
Notice that f4 could also have been described by a formula:
f4 .P; Q/ WWD ŒP IMPLIES Q:
A function might also be defined by a procedure for computing its value at any
element of its domain, or by some other kind of specification. For example, define
“mcs” — 2017/6/5 — 19:42 — page 104 — #112
104
Chapter 4
Mathematical Data Types
f5 .y/ to be the length of a left to right search of the bits in the binary string y until
a 1 appears, so
f5 .0010/ D 3;
f5 .100/ D 1;
f5 .0000/ is undefined:
Notice that f5 does not assign a value to any string of just 0’s. This illustrates
an important fact about functions: they need not assign a value to every element
in the domain. In fact this came up in our first example f1 .x/ D 1=x 2 , which
does not assign a value to 0. So in general, functions may be partial functions,
meaning that there may be domain elements for which the function is not defined.
The set of domain elements for which a function is defined is called the support of
the function. If a function assigns a value to every element of its domain, that is, its
support equals its domain, it is called a total function.
It’s often useful to find the set of values a function takes when applied to the
elements in a set of arguments. So if f W A ! B, and S is a subset of A, we define
f .S / to be the set of all the values that f takes when it is applied to elements of S .
That is,
f .S / WWD fb 2 B j f .s/ D b for some s 2 Sg:
For example, if we let Œr; s denote set of numbers in the interval from r to s on the
real line, then f1 .Œ1; 2/ D Œ1=4; 1.
For another example, let’s take the “search for a 1” function f5 . If we let X be
the set of binary words which start with an even number of 0’s followed by a 1,
then f5 .X / would be the odd nonnegative integers.
Applying f to a set S of arguments is referred to as “applying f pointwise to
S ”, and the set f .S / is referred to as the image of S under f .4 The set of values
that arise from applying f to all possible arguments is called the range of f . That
is,
range.f / WWD f .domain.f //:
Some authors refer to the codomain as the range of a function, but they shouldn’t.
The distinction between the range and codomain will be important later in Sections 4.5 when we relate sizes of sets to properties of functions between them.
4 There
is a picky distinction between the function f which applies to elements of A and the
function which applies f pointwise to subsets of A, because the domain of f is A, while the domain
of pointwise-f is pow.A/. It is usually clear from context whether f or pointwise-f is meant, so
there is no harm in overloading the symbol f in this way.
“mcs” — 2017/6/5 — 19:42 — page 105 — #113
4.4. Binary Relations
4.3.2
105
Function Composition
Doing things step by step is a universal idea. Taking a walk is a literal example, but
so is cooking from a recipe, executing a computer program, evaluating a formula,
and recovering from substance abuse.
Abstractly, taking a step amounts to applying a function, and going step by step
corresponds to applying functions one after the other. This is captured by the operation of composing functions. Composing the functions f and g means that first
f is applied to some argument, x, to produce f .x/, and then g is applied to that
result to produce g.f .x//.
Definition 4.3.1. For functions f W A ! B and g W B ! C , the composition,
g ı f , of g with f is defined to be the function from A to C defined by the rule:
.g ı f /.x/ WWD g.f .x//;
for all x 2 A.
Function composition is familiar as a basic concept from elementary calculus,
and it plays an equally basic role in discrete mathematics.
4.4
Binary Relations
Binary relations define relations between two objects. For example, “less-than” on
the real numbers relates every real number a to a real number b, precisely when
a < b. Similarly, the subset relation relates a set A to another set B precisely when
A B. A function f W A ! B is a special case of binary relation in which an
element a 2 A is related to an element b 2 B precisely when b D f .a/.
In this section we’ll define some basic vocabulary and properties of binary relations.
Definition 4.4.1. A binary relation R consists of a set A, called the domain of R,
a set B called the codomain of R, and a subset of A B called the graph of R.
A relation whose domain is A and codomain is B is said to be “between A and
B”, or “from A to B.” As with functions, we write R W A ! B to indicate that R
is a relation from A to B. When the domain and codomain are the same set A we
simply say the relation is “on A.” It’s common to use “a R b” to mean that the pair
.a; b/ is in the graph of R.5
5 Writing the relation or operator symbol between its arguments is called infix notation. Infix
expressions like “m < n” or “m C n” are the usual notation used for things like the less-then relation
or the addition operation rather than prefix notation like “< .m; n/” or “C.m; n/.”
“mcs” — 2017/6/5 — 19:42 — page 106 — #114
106
Chapter 4
Mathematical Data Types
Notice that Definition 4.4.1 is exactly the same as the definition in Section 4.3
of a function, except that it doesn’t require the functional condition that, for each
domain element a, there is at most one pair in the graph whose first coordinate is
a. As we said, a function is a special case of a binary relation.
The “in-charge of” relation Chrg for MIT in Spring ’10 subjects and instructors
is a handy example of a binary relation. Its domain Fac is the names of all the
MIT faculty and instructional staff, and its codomain is the set SubNums of subject
numbers in the Fall ’09–Spring ’10 MIT subject listing. The graph of Chrg contains
precisely the pairs of the form
.hinstructor-namei ; hsubject-numi/
such that the faculty member named hinstructor-namei is in charge of the subject
with number hsubject-numi that was offered in Spring ’10. So graph.Chrg/ contains pairs like
.T. Eng;
6.UAT/
.G. Freeman;
6.011/
.G. Freeman;
6.UAT/
.G. Freeman;
6.881/
.G. Freeman;
6.882/
.J. Guttag;
6.00/
(4.4)
.A. R. Meyer; 6.042/
.A. R. Meyer; 18.062/
.A. R. Meyer; 6.844/
.T. Leighton; 6.042/
.T. Leighton; 18.062/
::
:
Some subjects in the codomain SubNums do not appear among this list of pairs—
that is, they are not in range.Chrg/. These are the Fall term-only subjects. Similarly, there are instructors in the domain Fac who do not appear in the list because
they are not in charge of any Spring term subjects.
4.4.1
Relation Diagrams
Some standard properties of a relation can be visualized in terms of a diagram. The
diagram for a binary relation R has points corresponding to the elements of the
domain appearing in one column (a very long column if domain.R/ is infinite). All
the elements of the codomain appear in another column which we’ll usually picture
as being to the right of the domain column. There is an arrow going from a point a
in the left-hand, domain column to a point b in the right-hand, codomain column,
“mcs” — 2017/6/5 — 19:42 — page 107 — #115
4.4. Binary Relations
107
precisely when the corresponding elements are related by R. For example, here are
diagrams for two functions:
A
a
b PP
-
3
PP
Pq
c PP P
3
PP
Pq
d P
e B
A
1
a
-
1
2
b PP
3
PP
Pq
c Q P
2
3
4
B
Q
d QQ
3
4
QQ
s
5
Being a function is certainly an important property of a binary relation. What it
means is that every point in the domain column has at most one arrow coming out
of it. So we can describe being a function as the “ 1 arrow out” property. There
are four more standard properties of relations that come up all the time. Here are
all five properties defined in terms of arrows:
Definition 4.4.2. A binary relation R is:
a function when it has the Œ 1 arrow out property.
surjective when it has the Œ 1 arrows in property. That is, every point in
the right-hand, codomain column has at least one arrow pointing to it.
total when it has the Œ 1 arrows out property.
injective when it has the Œ 1 arrow in property.
bijective when it has both the ŒD 1 arrow out and the ŒD 1 arrow in property.
From here on, we’ll stop mentioning the arrows in these properties and for example, just write Œ 1 in instead of Œ 1 arrows in.
So in the diagrams above, the relation on the left has the ŒD 1 out and Œ 1 in
properties, which means it is a total, surjective function. But it does not have the
Œ 1 in property because element 3 has two arrows going into it; it is not injective.
The relation on the right has the ŒD 1 out and Œ 1 in properties, which means
it is a total, injective function. But it does not have the Œ 1 in property because
element 4 has no arrow going into it; it is not surjective.
The arrows in a diagram for R correspond, of course, exactly to the pairs in the
graph of R. Notice that the arrows alone are not enough to determine, for example,
“mcs” — 2017/6/5 — 19:42 — page 108 — #116
108
Chapter 4
Mathematical Data Types
if R has the Œ 1 out, total, property. If all we knew were the arrows, we wouldn’t
know about any points in the domain column that had no arrows out. In other
words, graph.R/ alone does not determine whether R is total: we also need to
know what domain.R/ is.
Example 4.4.3. The function defined by the formula 1=x 2 has the Œ 1 out property if its domain is RC , but not if its domain is some set of real numbers including
0. It has the ŒD 1 in and ŒD 1 out property if its domain and codomain are both
RC , but it has neither the Œ 1 in nor the Œ 1 out property if its domain and
codomain are both R.
4.4.2
Relational Images
The idea of the image of a set under a function extends directly to relations.
Definition 4.4.4. The image of a set Y under a relation R written R.Y /, is the set
of elements of the codomain B of R that are related to some element in Y . In terms
of the relation diagram, R.Y / is the set of points with an arrow coming in that starts
from some point in Y . The range range.R/ of R is the image R.A/ of the domain
A of R. That is, range.R/ is the set of all points in the codomain with an arrow
coming in.
For example, the set of subject numbers that Meyer is in charge of in Spring ’10
is exactly Chrg.A. Meyer/. To figure out what this is, we look for all the arrows
in the Chrg diagram that start at “A. Meyer,” and see which subject-numbers are
at the other end of these arrows. Looking at the list (4.4) of pairs in graph.Chrg/,
we see that these subject-numbers are f6.042, 18.062, 6.844g. Similarly, to find the
subject numbers that either Freeman or Eng are in charge of, we can collect all the
arrows that start at either “G. Freeman,” or “T. Eng” and, again, see which subjectnumbers are at the other end of these arrows. This is Chrg.fG. Freeman; T. Engg/.
Looking again at the list (4.4), we see that
Chrg.fG. Freeman; T. Engg/ D f6.011, 6.881, 6.882, 6.UATg
Finally, Fac is the set of all in-charge instructors, so Chrg.Fac/ is the set of all the
subjects listed for Spring ’10.
Inverse Relations and Images
Definition 4.4.5. The inverse, R 1 of a relation R W A ! B is the relation from B
to A defined by the rule
b R 1 a IFF a R b:
“mcs” — 2017/6/5 — 19:42 — page 109 — #117
4.5. Finite Cardinality
109
In other words, R 1 is the relation you get by reversing the direction of the
arrows in the diagram of R.
Definition 4.4.6. The inverse image of a set X B under the relation R is defined
to be R 1 .X /, namely, the set of elements in A connected by an arrow to some
element in B. The support support.R/ is defined to R 1 .B/, namely, the set of
domain elements with at least one arrow out. The support of R is also called the
domain of definition of R.
Continuing with the in-charge example above, the set of instructors in charge
of 6.UAT in Spring ’10 is exactly the inverse image of f6.UATg under the Chrg
relation. From the list (4.4), we see that Eng and Freeman are both in charge of
6.UAT, that is,
fT. Eng; D. Freemang Chrg
1
.f6.UATg/:
We can’t assert equality here because there may be additional pairs further down
the list showing that additional instructors are co-incharge of 6.UAT.
Now let Intro be the set of introductory course 6 subject numbers. These are the
subject numbers that start with “6.0.” So the set of names of the instructors who
were in-charge of introductory course 6 subjects in Spring ’10, is Chrg 1 .Intro/.
From the part of the Chrg list shown in (4.4), we see that Meyer, Leighton, Freeman, and Guttag were among the instructors in charge of introductory subjects in
Spring ’10. That is,
fMeyer, Leighton, Freeman, Guttagg Chrg
1
.Intro/:
Finally, Chrg 1 .SubNums/ is the set of all instructors who were in charge of a
subject listed for Spring ’10.
4.5
Finite Cardinality
A finite set is one that has only a finite number of elements. This number of elements is the “size” or cardinality of the set:
Definition 4.5.1. If A is a finite set, the cardinality jAj of A is the number of
elements in A.
A finite set may have no elements (the empty set), or one element, or two elements,. . . , so the cardinality of finite sets is always a nonnegative integer.
“mcs” — 2017/6/5 — 19:42 — page 110 — #118
110
Chapter 4
Mathematical Data Types
Now suppose R W A ! B is a function. This means that every element of A
contributes at most one arrow to the diagram for R, so the number of arrows is at
most the number of elements in A. That is, if R is a function, then
jAj #arrows:
If R is also surjective, then every element of B has an arrow into it, so there must
be at least as many arrows in the diagram as the size of B. That is,
#arrows jBj:
Combining these inequalities implies that if R is a surjective function, then jAj jBj.
In short, if we write A surj B to mean that there is a surjective function from
A to B, then we’ve just proved a lemma: if A surj B for finite sets A; B, then
jAj jBj. The following definition and lemma lists this statement and three similar
rules relating domain and codomain size to relational properties.
Definition 4.5.2. Let A; B be (not necessarily finite) sets. Then
1. A surj B iff there is a surjective function from A to B.
2. A inj B iff there is an injective total relation from A to B.
3. A bij B iff there is a bijection from A to B.
Lemma 4.5.3. For finite sets A; B:
1. If A surj B, then jAj jBj.
2. If A inj B, then jAj jBj.
3. If A bij B, then jAj D jBj.
Proof. We’ve already given an “arrow” proof of implication 1. Implication 2. follows immediately from the fact that if R has the Œ 1 out, function property, and
the Œ 1 in, surjective property, then R 1 is total and injective, so A surj B iff
B inj A. Finally, since a bijection is both a surjective function and a total injective
relation, implication 3. is an immediate consequence of the first two.
Lemma 4.5.3.1. has a converse: if the size of a finite set A is greater than or equal
to the size of another finite set B then it’s always possible to define a surjective
“mcs” — 2017/6/5 — 19:42 — page 111 — #119
4.5. Finite Cardinality
111
function from A to B. In fact, the surjection can be a total function. To see how
this works, suppose for example that
A D fa0 ; a1 ; a2 ; a3 ; a4 ; a5 g
B D fb0 ; b1 ; b2 ; b3 g:
Then define a total function f W A ! B by the rules
f .a0 / WWD b0 ; f .a1 / WWD b1 ; f .a2 / WWD b2 ; f .a3 / D f .a4 / D f .a5 / WWD b3 :
More concisely,
f .ai / WWD bmin.i;3/ ;
for 0 i 5. Since 5 3, this f is a surjection.
So we have figured out that if A and B are finite sets, then jAj jBj if and only if
A surj B. All told, this argument wraps up the proof of a theorem that summarizes
the whole finite cardinality story:
Theorem 4.5.4. [Mapping Rules] For finite sets A; B,
4.5.1
jAj jBj iff
A surj B;
(4.5)
jAj jBj iff
A inj B;
(4.6)
jAj D jBj iff
A bij B;
(4.7)
How Many Subsets of a Finite Set?
As an application of the bijection mapping rule (4.7), we can give an easy proof of:
Theorem 4.5.5. There are 2n subsets of an n-element set. That is,
jAj D n implies j pow.A/j D 2n :
For example, the three-element set fa1 ; a2 ; a3 g has eight different subsets:
;
fa1 g
fa2 g
fa1 ; a2 g
fa3 g fa1 ; a3 g fa2 ; a3 g fa1 ; a2 ; a3 g
Theorem 4.5.5 follows from the fact that there is a simple bijection from subsets
of A to f0; 1gn , the n-bit sequences. Namely, let a1 ; a2 ; : : : ; an be the elements
of A. The bijection maps each subset of S A to the bit sequence .b1 ; : : : ; bn /
defined by the rule that
bi D 1 iff ai 2 S:
“mcs” — 2017/6/5 — 19:42 — page 112 — #120
112
Chapter 4
Mathematical Data Types
For example, if n D 10, then the subset fa2 ; a3 ; a5 ; a7 ; a10 g maps to a 10-bit
sequence as follows:
subset: f
a2 ; a 3 ;
a5 ;
a7 ;
a10 g
sequence: . 0; 1; 1; 0; 1; 0; 1; 0; 0;
1 /
Now by bijection case of the Mapping Rules 4.5.4.(4.7),
j pow.A/j D jf0; 1gn j:
But every computer scientist knows6 that there are 2n n-bit sequences! So we’ve
proved Theorem 4.5.5!
Problems for Section 4.1
Practice Problems
Problem 4.1.
For any set A, let pow.A/ be its power set, the set of all its subsets; note that A is
itself a member of pow.A/. Let ; denote the empty set.
(a) The elements of pow.f1; 2g/ are:
(b) The elements of pow.f;; f;gg/ are:
(c) How many elements are there in pow.f1; 2; : : : ; 8g/?
Problem 4.2.
Express each of the following assertions about sets by a formula of set theory.7
Expressions may use abbreviations introduced earlier (so it is now legal to use “D”
because we just defined it).
(a) x D ;.
(b) x D fy; zg.
(c) x y. (x is a subset of y that might equal y.)
6 In
2n
case you’re someone who doesn’t know how many n-bit sequences there are, you’ll find the
explained in Section 15.2.2.
7 See Section 8.3.2.
“mcs” — 2017/6/5 — 19:42 — page 113 — #121
4.5. Finite Cardinality
113
Now we can explain how to express “x is a proper subset of y” as a set theory
formula using things we already know how to express. Namely, letting “x ¤ y”
abbreviate NOT.x D y/, the expression
.x y AND x ¤ y/;
describes a formula of set theory that means x y.
From here on, feel free to use any previously expressed property in describing
formulas for the following:
(d) x D y [ z.
(e) x D y
z.
(f) x D pow.y/.
S
(g) x D z2y z.
This means that y is supposed to be S
a collection of sets, and
S x is the union of all of
them. A more concise notation for “ z2y z’ is simply “ y.”
Class Problems
Problem 4.3.
Set Formulas and Propositional Formulas.
(a) Verify that the propositional formula .P AND Q/ OR .P AND Q/ is equivalent
to P .
(b) Prove that
A D .A
B/ [ .A \ B/
for all sets, A; B, by showing
x 2 A IFF x 2 .A
B/ [ .A \ B/
for all elements x using the equivalence of part (a) in a chain of IFF’s.
Problem 4.4.
Prove
Theorem (Distributivity of union over intersection).
A [ .B \ C / D .A [ B/ \ .A [ C /
(4.8)
“mcs” — 2017/6/5 — 19:42 — page 114 — #122
114
Chapter 4
Mathematical Data Types
for all sets, A; B; C , by using a chain of iff’s to show that
x 2 A [ .B \ C / IFF x 2 .A [ B/ \ .A [ C /
for all elements x. You may assume the corresponding propositional equivalence 3.10.
Problem 4.5.
Prove De Morgan’s Law for set equality
A \ B D A [ B:
(4.9)
by showing with a chain of IFF’s that x 2 the left-hand side of (4.9) iff x 2 the
right-hand side. You may assume the propositional version (3.14) of De Morgan’s
Law.
Problem 4.6.
Powerset Properties.
Let A and B be sets.
(a) Prove that
pow.A \ B/ D pow.A/ \ pow.B/:
(b) Prove that
.pow.A/ [ pow.B// pow.A [ B/;
with equality holding iff one of A or B is a subset of the other.
Problem 4.7.
Subset take-away8 is a two player game played with a finite set A of numbers.
Players alternately choose nonempty subsets of A with the conditions that a player
may not choose
the whole set A, or
any set containing a set that was named earlier.
8 From Christenson & Tilford, David Gale’s Subset Takeaway Game, American Mathematical
Monthly, Oct. 1997
“mcs” — 2017/6/5 — 19:42 — page 115 — #123
4.5. Finite Cardinality
115
The first player who is unable to move loses the game.
For example, if the size of A is one, then there are no legal moves and the second
player wins. If A has exactly two elements, then the only legal moves are the two
one-element subsets of A. Each is a good reply to the other, and so once again the
second player wins.
The first interesting case is when A has three elements. This time, if the first
player picks a subset with one element, the second player picks the subset with the
other two elements. If the first player picks a subset with two elements, the second
player picks the subset whose sole member is the third element. In both cases, these
moves lead to a situation that is the same as the start of a game on a set with two
elements, and thus leads to a win for the second player.
Verify that when A has four elements, the second player still has a winning strategy.9
Homework Problems
Problem 4.8.
Let A, B and C be sets. Prove that
A [ B [ C D .A
B/ [ .B
C / [ .C
A/ [ .A \ B \ C /
(4.10)
using a chain of IFF’s as Section 4.1.5.
Problem 4.9.
Union distributes over the intersection of two sets:
A [ .B \ C / D .A [ B/ \ .A [ C /
(4.11)
(see Problem 4.4).
Use (4.11) and the Well Ordering Principle to prove the Distributive Law of
union over the intersection of n sets:
A [ .B1 \ \ Bn
1
\ Bn /
D .A [ B1 / \ \ .A [ Bn
1/
\ .A [ Bn /
(4.12)
Extending formulas to an arbitrary number of terms is a common (if mundane)
application of the WOP.
9 David
Gale worked out some of the properties of this game and conjectured that the second
player wins the game for any set A. This remains an open problem.
“mcs” — 2017/6/5 — 19:42 — page 116 — #124
116
Chapter 4
Mathematical Data Types
Exam Problems
Problem 4.10.
You’ve seen how certain set identities follow from corresponding propositional
equivalences. For example, you proved by a chain of iff’s that
.A
B/ [ .A \ B/ D A
using the fact that the propositional formula .P AND Q/ OR .P AND Q/ is equivalent
to P .
State a similar propositional equivalence that would justify the key step in a proof
for the following set equality organized as a chain of iff’s:
A B D A C [ .B \ C / [ A [ B \ C
(4.13)
(You are not being asked to write out an iff-proof of the equality or to write out
a proof of the propositional equivalence. Just state the equivalence.)
Problem 4.11.
You’ve seen how certain set identities follow from corresponding propositional
equivalences. For example, you proved by a chain of iff’s that
.A
B/ [ .A \ B/ D A
using the fact that the propositional formula .P AND Q/ OR .P AND Q/ is equivalent
to P .
State a similar propositional equivalence that would justify the key step in a proof
for the following set equality organized as a chain of iff’s:
A \ B \ C D A [ .B
A/ [ C :
(You are not being asked to write out an iff-proof of the equality or to write out
a proof of the propositional equivalence. Just state the equivalence.)
Problem 4.12.
The set equation
A\B DA[B
follows from a certain equivalence between propositional formulas.
(a) What is the equivalence?
(b) Show how to derive the equation from this equivalence.
“mcs” — 2017/6/5 — 19:42 — page 117 — #125
4.5. Finite Cardinality
117
Problems for Section 4.2
Homework Problems
Problem 4.13.
Prove that for any sets A, B, C and D, if the Cartesian products A B and C D
are disjoint, then either A and C are disjoint or B and D are disjoint.
Problem 4.14. (a) Give a simple example where the following result fails, and
briefly explain why:
False Theorem. For sets A, B, C and D, let
L WWD .A [ B/ .C [ D/;
R WWD .A C / [ .B D/:
Then L D R.
(b) Identify the mistake in the following proof of the False Theorem.
Bogus proof. Since L and R are both sets of pairs, it’s sufficient to prove that
.x; y/ 2 L ! .x; y/ 2 R for all x; y.
The proof will be a chain of iff implications:
iff
iff
iff
iff
iff
iff
.x; y/ 2 R
.x; y/ 2 .A C / [ .B D/
.x; y/ 2 A C , or .x; y/ 2 B D
(x 2 A and y 2 C ) or else (x 2 B and y 2 D)
either x 2 A or x 2 B, and either y 2 C or y 2 D
x 2 A [ B and y 2 C [ D
.x; y/ 2 L.
(c) Fix the proof to show that R L.
“mcs” — 2017/6/5 — 19:42 — page 118 — #126
118
Chapter 4
Mathematical Data Types
Problems for Section 4.4
Practice Problems
Problem 4.15.
The inverse R
defined by:
1
of a binary relation R from A to B is the relation from B to A
bR
1
a
iff a R b:
In other words, you get the diagram for R 1 from R by “reversing the arrows” in
the diagram describing R. Now many of the relational properties of R correspond
to different properties of R 1 . For example, R is total iff R 1 is a surjection.
Fill in the remaining entries is this table:
R is
total
a function
a surjection
an injection
a bijection
iff
R 1 is
a surjection
Hint: Explain what’s going on in terms of “arrows” from A to B in the diagram
for R.
Problem 4.16.
Describe a total injective function ŒD 1 out, Œ 1 in;  from R ! R that is not a
bijection.
Problem 4.17.
For a binary relation R W A ! B, some properties of R can be determined from
just the arrows of R, that is, from graph.R/, and others require knowing if there
are elements in the domain A or the codomain B that don’t show up in graph.R/.
For each of the following possible properties of R, indicate whether it is always
determined by
1. graph.R/ alone,
2. graph.R/ and A alone,
3. graph.R/ and B alone,
“mcs” — 2017/6/5 — 19:42 — page 119 — #127
4.5. Finite Cardinality
119
4. all three parts of R.
Properties:
(a) surjective
(b) injective
(c) total
(d) function
(e) bijection
Problem 4.18.
For each of the following real-valued functions on the real numbers, indicate whether
it is a bijection, a surjection but not a bijection, an injection but not a bijection, or
neither an injection nor a surjection.
(a) x ! x C 2
(b) x ! 2x
(c) x ! x 2
(d) x ! x 3
(e) x ! sin x
(f) x ! x sin x
(g) x ! e x
Problem 4.19.
Let f W A ! B and g W B ! C be functions and h W A ! C be their composition,
namely, h.a/ WWD g.f .a// for all a 2 A.
(a) Prove that if f and g are surjections, then so is h.
(b) Prove that if f and g are bijections, then so is h.
(c) If f is a bijection, then so is f
1.
“mcs” — 2017/6/5 — 19:42 — page 120 — #128
120
Chapter 4
Mathematical Data Types
Problem 4.20.
Give an example of a relation R that is a total injective function from a set A to
itself but is not a bijection.
Class Problems
Problem 4.21. (a) Prove that if A surj B and B surj C , then A surj C .
(b) Explain why A surj B iff B inj A.
(c) Conclude from (a) and (b) that if A inj B and B inj C , then A inj C .
(d) According to Definition 4.5.2, A inj B requires a total injective relation. Explain why A inj B iff there is a total injective function from A to B.
Problem 4.22.
Five basic properties of binary relations R W A ! B are:
1. R is a surjection Œ 1 in
2. R is an injection Œ 1 in
3. R is a function Œ 1 out
4. R is total Œ 1 out
5. R is empty ŒD 0 out
Below are some assertions about R. For each assertion, indicate all the properties
above that the relation R must have. For example, the first assertion impllies that R
is a total surjection. Variables a; a1 ; : : : range over A and b; b1 ; : : : range over B.
(a) 8a 8b: a R b.
(b) NOT.8a 8b: a R b/.
(c) 8a 8b: NOT.a R b/.
(d) 8a 9b: a R b.
(e) 8b 9a: a R b.
(f) R is a bijection.
V
(g) 8a 9b1 a R b1 8b: a R b IMPLIES b D b1 .
“mcs” — 2017/6/5 — 19:42 — page 121 — #129
4.5. Finite Cardinality
121
(h) 8a; b: a R b OR a ¤ b.
(i) 8b1 ; b2 ; a: .a R b1 AND a R b2 / IMPLIES b1 D b2 .
(j) 8a1 ; a2 ; b: .a1 R b AND a2 R b/ IMPLIES a1 D a2 .
(k) 8a1 ; a2 ; b1 ; b2 : .a1 R b1 AND a2 R b2 AND a1 ¤ a2 / IMPLIES b1 ¤ b2 .
(l) 8a1 ; a2 ; b1 ; b2 : .a1 R b1 AND a2 R b2 AND b1 ¤ b2 / IMPLIES a1 ¤ a2 .
Problem 4.23.
Let R W A ! B be a binary relation. Each of the following formulas expresses
the fact that R has a familiar relational “arrow” property such as being surjective
or being a function.
Identify the relational property expressed by each of the following relational
expressions. Explain your reasoning.
(a) R ı R
1
IdB
(b) R
1
ı R IdA
(c) R
1
ı R IdA
(d) R ı R
1
IdB
Homework Problems
Problem 4.24.
Let f W A ! B and g W B ! C be functions.
(a) Prove that if the composition g ı f is a bijection, then f is a total injection
and g is a surjection.
(b) Show there is a total injection f and a bijection, g, such that g ı f is not a
bijection.
Problem 4.25.
Let A, B and C be nonempty sets, and let f W B ! C and g W A ! B be
functions. Let h WWD f ı g be the composition function of f and g, namely, the
function with domain A and codomain C such that h.x/ D f .g.x//.
(a) Prove that if h is surjective and f is total and injective, then g must be surjective.
“mcs” — 2017/6/5 — 19:42 — page 122 — #130
122
Chapter 4
Mathematical Data Types
Hint: contradiction.
(b) Suppose that h is injective and f is total. Prove that g must be injective and
provide a counterexample showing how this claim could fail if f was not total.
Problem 4.26.
Let A, B and C be sets, and let f W B ! C and g W A ! B be functions. Let
h W A ! C be the composition f ı g; that is, h.x/ WWD f .g.x// for x 2 A. Prove
or disprove the following claims:
(a) If h is surjective, then f must be surjective.
(b) If h is surjective, then g must be surjective.
(c) If h is injective, then f must be injective.
(d) If h is injective and f is total, then g must be injective.
Problem 4.27. (a)
Let R W D ! D be a binary relation on a set D. Let x; y be variables ranging over
D. Indicate the expressions below whose meaning is that R is an injective relation
Œ 1 in. Remember that R.x/ WWD fy j x R yg, and R is not necessarily a function
or a total relation.
(i) R.x/ \ R.y/ D ;
(ii) R.x/ D R.y/ IMPLIES x D y
(iii) R.x/ \ R.y/ D ; IMPLIES x ¤ y
(iv) x ¤ y IMPLIES R.x/ ¤ R.y/
(v) R.x/ \ R.y/ ¤ ; IMPLIES x ¤ y
(vi) R.x/ \ R.y/ ¤ ; IMPLIES x D y
(vii) R
1 .R.x//
D fxg
(viii) R
1 .R.x//
fxg
(ix) R
1 .R.x//
fxg
(x) R.R
1 .x//
fxg
(xi) R.R
1 .x//
fxg
(xii) x ¤ y IMPLIES R.x/ \ R.y/ D ;
“mcs” — 2017/6/5 — 19:42 — page 123 — #131
4.5. Finite Cardinality
123
(b) Give an example of a set S such that there is no total injective relation from S
to the real interval Œ0; 1.
Problem 4.28.
The language of sets and relations may seem remote from the practical world of
programming, but in fact there is a close connection to relational databases, a
very popular software application building block implemented by such software
packages as MySQL. This problem explores the connection by considering how to
manipulate and analyze a large data set using operators over sets and relations. Systems like MySQL are able to execute very similar high-level instructions efficiently
on standard computer hardware, which helps programmers focus on high-level design.
Consider a basic Web search engine, which stores information on Web pages and
processes queries to find pages satisfying conditions provided by users. At a high
level, we can formalize the key information as:
A set P of pages that the search engine knows about
A binary relation L (for link) over pages, defined such that p1 L p2 iff page
p1 links to p2
A set E of endorsers, people who have recorded their opinions about which
pages are high-quality
A binary relation R (for recommends) between endorsers and pages, such
that e R p iff person e has recommended page p
A set W of words that may appear on pages
A binary relation M (for mentions) between pages and words, where p M w
iff word w appears on page p
Each part of this problem describes an intuitive, informal query over the data,
and your job is to produce a single expression using the standard set and relation
operators, such that the expression can be interpreted as answering the query correctly, for any data set. Your answers should use only the set and relation symbols
given above, in addition to terms standing for constant elements of E or W , plus
the following operators introduced in the text:
set union [.
“mcs” — 2017/6/5 — 19:42 — page 124 — #132
124
Chapter 4
Mathematical Data Types
set intersection \.
set difference .
relational image—for example, R.A/ for some set A, or R.a/ for some specific element a.
relational inverse
1.
. . . and one extra: relational composition which generalizes composition of
functions
a .R ı S / c WWD 9b 2 B: .a S b/ AND .b R c/:
In other words, a is related to c in R ı S if starting at a you can follow an S
arrow to the start of an R arrow and then follow the R arrow to get to c.10
Here is one worked example to get you started:
Search description: The set of pages containing the word “logic”
Solution expression: M
1 .“logic”/
Find similar solutions for each of the following searches:
(a) The set of pages containing the word “logic” but not the word “predicate”
(b) The set of pages containing the word “set” that have been recommended by
“Meyer”
(c) The set of endorsers who have recommended pages containing the word “algebra”
(d) The relation that relates endorser e and word w iff e has recommended a page
containing w
(e) The set of pages that have at least one incoming or outgoing link
(f) The relation that relates word w and page p iff w appears on a page that links
to p
(g) The relation that relates word w and endorser e iff w appears on a page that
links to a page that e recommends
(h) The relation that relates pages p1 and p2 iff p2 can be reached from p1 by
following a sequence of exactly 3 links
10 Note the reversal of R and S in the definition; this is to make relational composition work like
function composition. For functions, f ı g means you apply g first. That is, if we let h be f ı g,
then h.x/ D f .g.x//.
“mcs” — 2017/6/5 — 19:42 — page 125 — #133
4.5. Finite Cardinality
125
Exam Problems
Problem 4.29.
Let A be the set containing the five sets: fag; fb; cg; fb; d g; fa; eg; fe; f g, and let
B be the set containing the three sets: fa; bg; fb; c; d g; fe; f g. Let R be the “is
subset of” binary relation from A to B defined by the rule:
XRY
IFF
X Y:
(a) Fill in the arrows so the following figure describes the graph of the relation,
R:
A
B
arrows
fag
fa; bg
fb; cg
fb; c; d g
fb; d g
fe; f g
fa; eg
fe; f g
(b) Circle the properties below possessed by the relation R:
function
total
injective
surjective
(c) Circle the properties below possessed by the relation R
function
total
injective
surjective
bijective
1:
bijective
“mcs” — 2017/6/5 — 19:42 — page 126 — #134
126
Chapter 4
Mathematical Data Types
Problem 4.30. (a) Five assertions about a binary relation R W A ! B are bulleted
below. There are nine predicate formulas that express some of these assertions.
Write the numbers of the formulas next to the assertions they express. For example,
you should write “4” next to the last assertion, since formula (4) expresses the
assertion that R is the identity relation.
Variables a; a1 ; : : : range over the domain A and b; b1 ; : : : range over the codomain
B. More than one formula may express one assertion.
R is a surjection
R is an injection
R is a function
R is total
R is the identity relation.
1. 8b: 9a: a R b.
2. 8a: 9b: a R b.
3. 8a: a R a.
4. 8a; b: a R b IFF a D b.
5. 8a; b: a R b OR a ¤ b.
6. 8b1 ; b2 ; a: .a R b1 AND a R b2 / IMPLIES b1 D b2 .
7. 8a1 ; a2 ; b: .a1 R b AND a2 R b/ IMPLIES a1 D a2 .
8. 8a1 ; a2 ; b1 ; b2 : .a1 R b1 AND a2 R b2 AND a1 ¤ a2 / IMPLIES b1 ¤ b2 .
9. 8a1 ; a2 ; b1 ; b2 : .a1 R b1 AND a2 R b2 AND b1 ¤ b2 / IMPLIES a1 ¤ a2 .
(b) Give an example of a relation R that satisfies three of the properties surjection,
injection, total, and function (you indicate which) but is not a bijection.
Problem 4.31.
Let f W D ! D be a total function from some nonempty set D to itself. In the
following propositions, x and y are variables ranging over D, and g is a variable
ranging over total functions from D to D. Indicate all of the propositions that are
equivalent to the proposition that f is an injection:
1. x D y OR f .x/ ¤ f .y/
2. x D y IMPLIES f .x/ D f .y/
“mcs” — 2017/6/5 — 19:42 — page 127 — #135
4.5. Finite Cardinality
127
3. x ¤ y IMPLIES f .x/ ¤ f .y/
4. f .x/ D f .y/ IMPLIES x D y
5. NOTŒ9x9y.x ¤ y AND f .x/ D f .y//
6. NOTŒ9z8x.f .x/ ¤ z/
7. 9g8x.g.f .x// D x/
8. 9g8x.f .g.x// D x/
Problem 4.32.
Prove that if relation R W A ! B is a total injection, Œ 1 out; Œ 1 in, then
R
1
ı R D IdA ;
where IdA is the identity function on A.
(A simple argument in terms of ”arrows” will do the job.)
Problem 4.33.
Let R W A ! B be a binary relation.
(a) Prove that R is a function iff R ı R 1 IdB .
Write similar containment formulas involving R 1 ıR, RıR 1 , Ida , IdB equivalent
to the assertion that R has each of the following properties. No proof is required.
(b) total.
(c) a surjection.
(d) a injection.
Problem 4.34.
Let R W A ! B and S W B ! C be binary relations such that S ı R is a bijection
and jAj D 2.
Give an example of such R; S where neither R nor S is a function. Indicate exactly which properties—total, surjection, function, and injection—your examples
of R and S have.
Hint: Let jBj D 4.
“mcs” — 2017/6/5 — 19:42 — page 128 — #136
128
Chapter 4
Mathematical Data Types
Problem 4.35.
The set f1; 2; 3g! consists of the infinite sequences of the digits 1,2, and 3, and
likewise f4; 5g! is the set of infinite sequences of the digits 4,5. For example
123123123 : : :
222222222222 : : :
4554445554444 : : :
2 f1; 2; 3g! ;
2 f1; 2; 3g! ;
2 f4; 5g! :
(a) Give an example of a total injective function
f W f1; 2; 3g! ! f4; 5g! :
(b) Give an example of a bijection g W .f1; 2; 3g! f1; 2; 3g! / ! f1; 2; 3g! .
(c) Explain why there is a bijection between f1; 2; 3g! f1; 2; 3g! and f4; 5g! .
(You need not explicitly define the bijection.)
Problems for Section 4.5
Practice Problems
Problem 4.36.
Assume f W A ! B is total function, and A is finite. Replace the ? with one of
; D; to produce the strongest correct version of the following statements:
(a) jf .A/j ? jBj.
(b) If f is a surjection, then jAj ? jBj.
(c) If f is a surjection, then jf .A/j ? jBj.
(d) If f is an injection, then jf .A/j ? jAj.
(e) If f is a bijection, then jAj ? jBj.
Class Problems
Problem 4.37.
Let A D fa0 ; a1 ; : : : ; an 1 g be a set of size n, and B D fb0 ; b1 ; : : : ; bm 1 g a set
of size m. Prove that jA Bj D mn by defining a simple bijection from A B to
the nonnegative integers from 0 to mn 1.
“mcs” — 2017/6/5 — 19:42 — page 129 — #137
4.5. Finite Cardinality
129
Problem 4.38.
Let R W A ! B be a binary relation. Use an arrow counting argument to prove the
following generalization of the Mapping Rule 1.
Lemma. If R is a function, and X A, then
jXj jR.X /j:
“mcs” — 2017/6/5 — 19:42 — page 130 — #138
“mcs” — 2017/6/5 — 19:42 — page 131 — #139
5
Induction
Induction is a powerful method for showing a property is true for all nonnegative
integers. Induction plays a central role in discrete mathematics and computer science. In fact, its use is a defining characteristic of discrete—as opposed to continuous—mathematics. This chapter introduces two versions of induction, Ordinary
and Strong, and explains why they work and how to use them in proofs. It also
introduces the Invariant Principle, which is a version of induction specially adapted
for reasoning about step-by-step processes.
5.1
Ordinary Induction
To understand how induction works, suppose there is a professor who brings a
bottomless bag of assorted miniature candy bars to her large class. She offers to
share the candy in the following way. First, she lines the students up in order. Next
she states two rules:
1. The student at the beginning of the line gets a candy bar.
2. If a student gets a candy bar, then the following student in line also gets a
candy bar.
Let’s number the students by their order in line, starting the count with 0, as usual
in computer science. Now we can understand the second rule as a short description
of a whole sequence of statements:
If student 0 gets a candy bar, then student 1 also gets one.
If student 1 gets a candy bar, then student 2 also gets one.
If student 2 gets a candy bar, then student 3 also gets one.
::
:
Of course, this sequence has a more concise mathematical description:
If student n gets a candy bar, then student n C 1 gets a candy bar, for
all nonnegative integers n.
“mcs” — 2017/6/5 — 19:42 — page 132 — #140
132
Chapter 5
Induction
So suppose you are student 17. By these rules, are you entitled to a miniature candy
bar? Well, student 0 gets a candy bar by the first rule. Therefore, by the second
rule, student 1 also gets one, which means student 2 gets one, which means student
3 gets one as well, and so on. By 17 applications of the professor’s second rule,
you get your candy bar! Of course the rules really guarantee a candy bar to every
student, no matter how far back in line they may be.
5.1.1
A Rule for Ordinary Induction
The reasoning that led us to conclude that every student gets a candy bar is essentially all there is to induction.
The Induction Principle.
Let P be a predicate on nonnegative integers. If
P .0/ is true, and
P .n/ IMPLIES P .n C 1/ for all nonnegative integers n
then
P .m/ is true for all nonnegative integers m.
Since we’re going to consider several useful variants of induction in later sections, we’ll refer to the induction method described above as ordinary induction
when we need to distinguish it. Formulated as a proof rule as in Section 1.4.1, this
would be
Rule. Induction Rule
P .0/;
8n 2 N: P .n/ IMPLIES P .n C 1/
8m 2 N: P .m/
This Induction Rule works for the same intuitive reason that all the students get
candy bars, and we hope the explanation using candy bars makes it clear why the
soundness of ordinary induction can be taken for granted. In fact, the rule is so
obvious that it’s hard to see what more basic principle could be used to justify it.1
What’s not so obvious is how much mileage we get by using it.
1 But
see Section 5.3.
“mcs” — 2017/6/5 — 19:42 — page 133 — #141
5.1. Ordinary Induction
5.1.2
133
A Familiar Example
Below is the formula (5.1) for the sum of the nonnegative integers up to n. The
formula holds for all nonnegative integers, so it is the kind of statement to which
induction applies directly. We’ve already proved this formula using the Well Ordering Principle (Theorem 2.2.1), but now we’ll prove it by induction, that is, using
the Induction Principle.
Theorem 5.1.1. For all n 2 N,
1 C 2 C 3 C C n D
n.n C 1/
2
(5.1)
To prove the theorem by induction, define predicate P .n/ to be the equation (5.1).
Now the theorem can be restated as the claim that P .n/ is true for all n 2 N. This
is great, because the Induction Principle lets us reach precisely that conclusion,
provided we establish two simpler facts:
P .0/ is true.
For all n 2 N, P .n/ IMPLIES P .n C 1/.
So now our job is reduced to proving these two statements.
The first statement follows because of the convention that a sum of zero terms
is equal to 0. So P .0/ is the true assertion that a sum of zero terms is equal to
0.0 C 1/=2 D 0.
The second statement is more complicated. But remember the basic plan from
Section 1.5 for proving the validity of any implication: assume the statement on
the left and then prove the statement on the right. In this case, we assume P .n/—
namely, equation (5.1)—in order to prove P .n C 1/, which is the equation
1 C 2 C 3 C C n C .n C 1/ D
.n C 1/.n C 2/
:
2
(5.2)
These two equations are quite similar; in fact, adding .n C 1/ to both sides of
equation (5.1) and simplifying the right side gives the equation (5.2):
n.n C 1/
C .n C 1/
2
.n C 2/.n C 1/
D
2
1 C 2 C 3 C C n C .n C 1/ D
Thus, if P .n/ is true, then so is P .n C 1/. This argument is valid for every nonnegative integer n, so this establishes the second fact required by the induction
proof. Therefore, the Induction Principle says that the predicate P .m/ is true for
all nonnegative integers m. The theorem is proved.
“mcs” — 2017/6/5 — 19:42 — page 134 — #142
134
Chapter 5
5.1.3
Induction
A Template for Induction Proofs
The proof of equation (5.1) was relatively simple, but even the most complicated
induction proof follows exactly the same template. There are five components:
1. State that the proof uses induction. This immediately conveys the overall
structure of the proof, which helps your reader follow your argument.
2. Define an appropriate predicate P .n/. The predicate P .n/ is called the
induction hypothesis. The eventual conclusion of the induction argument
will be that P .n/ is true for all nonnegative n. A clearly stated induction
hypothesis is often the most important part of an induction proof, and its
omission is the largest source of confused proofs by students.
In the simplest cases, the induction hypothesis can be lifted straight from the
proposition you are trying to prove, as we did with equation (5.1). Sometimes
the induction hypothesis will involve several variables, in which case you
should indicate which variable serves as n.
3. Prove that P .0/ is true. This is usually easy, as in the example above. This
part of the proof is called the base case or basis step.
4. Prove that P .n/ implies P .n C 1/ for every nonnegative integer n. This
is called the inductive step. The basic plan is always the same: assume that
P .n/ is true, and then use this assumption to prove that P .n C 1/ is true.
These two statements should be fairly similar, but bridging the gap may require some ingenuity. Whatever argument you give must be valid for every
nonnegative integer n, since the goal is to prove that all the following implications are true:
P .0/ ! P .1/; P .1/ ! P .2/; P .2/ ! P .3/; : : : :
5. Invoke induction. Given these facts, the induction principle allows you to
conclude that P .n/ is true for all nonnegative n. This is the logical capstone
to the whole argument, but it is so standard that it’s usual not to mention it
explicitly.
Always be sure to explicitly label the base case and the inductive step. Doing
so will make your proofs clearer and will decrease the chance that you forget a key
step—like checking the base case.
“mcs” — 2017/6/5 — 19:42 — page 135 — #143
5.1. Ordinary Induction
5.1.4
135
A Clean Writeup
The proof of Theorem 5.1.1 given above is perfectly valid; however, it contains a
lot of extraneous explanation that you won’t usually see in induction proofs. The
writeup below is closer to what you might see in print and should be prepared to
produce yourself.
Revised proof of Theorem 5.1.1. We use induction. The induction hypothesis P .n/
will be equation (5.1).
Base case: P .0/ is true, because both sides of equation (5.1) equal zero when
n D 0.
Inductive step: Assume that P .n/ is true, that is equation (5.1) holds for some
nonnegative integer n. Then adding n C 1 to both sides of the equation implies that
n.n C 1/
C .n C 1/
2
.n C 1/.n C 2/
D
2
1 C 2 C 3 C C n C .n C 1/ D
(by simple algebra)
which proves P .n C 1/.
So it follows by induction that P .n/ is true for all nonnegative n.
It probably bothers you that induction led to a proof of this summation formula
but did not provide an intuitive way to understand it, nor did it explain where the
formula came from in the first place.2 This is both a weakness and a strength. It is a
weakness when a proof does not provide insight. But it is a strength that a proof can
provide a reader with a reliable guarantee of correctness without requiring insight.
5.1.5
A More Challenging Example
During the development of MIT’s famous Stata Center, as costs rose further and
further beyond budget, some radical fundraising ideas were proposed. One rumored
plan was to install a big square courtyard divided into unit squares. The big square
would be 2n units on a side for some undetermined nonnegative integer n, and
one of the unit squares in the center3 occupied by a statue of a wealthy potential
donor—whom the fund raisers privately referred to as “Bill.” The n D 3 case is
shown in Figure 5.1.
A complication was that the building’s unconventional architect, Frank Gehry,
was alleged to require that only special L-shaped tiles (shown in Figure 5.2) be
2 Methods
for finding such formulas are covered in Part III of the text.
the special case n D 0, the whole courtyard consists of a single central square; otherwise,
there are four central squares.
3 In
“mcs” — 2017/6/5 — 19:42 — page 136 — #144
136
Chapter 5
Induction
2n
2n
Figure 5.1 A 2n 2n courtyard for n D 3.
Figure 5.2
The special L-shaped tile.
used for the courtyard. For n D 2, a courtyard meeting these constraints is shown
in Figure 5.3. But what about for larger values of n? Is there a way to tile a 2n 2n
courtyard with L-shaped tiles around a statue in the center? Let’s try to prove that
this is so.
Theorem 5.1.2. For all n 0 there exists a tiling of a 2n 2n courtyard with Bill
in a central square.
Proof. (doomed attempt) The proof is by induction. Let P .n/ be the proposition
that there exists a tiling of a 2n 2n courtyard with Bill in the center.
Base case: P .0/ is true because Bill fills the whole courtyard.
Inductive step: Assume that there is a tiling of a 2n 2n courtyard with Bill in the
center for some n 0. We must prove that there is a way to tile a 2nC1 2nC1
courtyard with Bill in the center . . . .
Now we’re in trouble! The ability to tile a smaller courtyard with Bill in the
“mcs” — 2017/6/5 — 19:42 — page 137 — #145
5.1. Ordinary Induction
137
B
Figure 5.3 A tiling using L-shaped tiles for n D 2 with Bill in a center square.
center isn’t much help in tiling a larger courtyard with Bill in the center. We haven’t
figured out how to bridge the gap between P .n/ and P .n C 1/.
So if we’re going to prove Theorem 5.1.2 by induction, we’re going to need some
other induction hypothesis than simply the statement about n that we’re trying to
prove.
When this happens, your first fallback should be to look for a stronger induction
hypothesis; that is, one which implies your previous hypothesis. For example,
we could make P .n/ the proposition that for every location of Bill in a 2n 2n
courtyard, there exists a tiling of the remainder.
This advice may sound bizarre: “If you can’t prove something, try to prove something grander!” But for induction arguments, this makes sense. In the inductive
step, where you have to prove P .n/ IMPLIES P .n C 1/, you’re in better shape
because you can assume P .n/, which is now a more powerful statement. Let’s see
how this plays out in the case of courtyard tiling.
Proof (successful attempt). The proof is by induction. Let P .n/ be the proposition
that for every location of Bill in a 2n 2n courtyard, there exists a tiling of the
remainder.
Base case: P .0/ is true because Bill fills the whole courtyard.
Inductive step: Assume that P .n/ is true for some n 0; that is, for every location
of Bill in a 2n 2n courtyard, there exists a tiling of the remainder. Divide the
2nC1 2nC1 courtyard into four quadrants, each 2n 2n . One quadrant contains
Bill (B in the diagram below). Place a temporary Bill (X in the diagram) in each of
the three central squares lying outside this quadrant as shown in Figure 5.4.
“mcs” — 2017/6/5 — 19:42 — page 138 — #146
138
Chapter 5
Induction
B
2n
X
X X
2n
2n
2n
Figure 5.4 Using a stronger inductive hypothesis to prove Theorem 5.1.2.
Now we can tile each of the four quadrants by the induction assumption. Replacing the three temporary Bills with a single L-shaped tile completes the job. This
proves that P .n/ implies P .n C 1/ for all n 0. Thus P .m/ is true for all m 2 N,
and the theorem follows as a special case where we put Bill in a central square. This proof has two nice properties. First, not only does the argument guarantee
that a tiling exists, but also it gives an algorithm for finding such a tiling. Second,
we have a stronger result: if Bill wanted a statue on the edge of the courtyard, away
from the pigeons, we could accommodate him!
Strengthening the induction hypothesis is often a good move when an induction
proof won’t go through. But keep in mind that the stronger assertion must actually
be true; otherwise, there isn’t much hope of constructing a valid proof. Sometimes
finding just the right induction hypothesis requires trial, error, and insight. For
example, mathematicians spent almost twenty years trying to prove or disprove
the conjecture that every planar graph is 5-choosable.4 Then, in 1994, Carsten
Thomassen gave an induction proof simple enough to explain on a napkin. The
key turned out to be finding an extremely clever induction hypothesis; with that in
hand, completing the argument was easy!
4 5-choosability
is a slight generalization of 5-colorability. Although every planar graph is 4colorable and therefore 5-colorable, not every planar graph is 4-choosable. If this all sounds like
nonsense, don’t panic. We’ll discuss graphs, planarity, and coloring in Part II of the text.
“mcs” — 2017/6/5 — 19:42 — page 139 — #147
5.1. Ordinary Induction
5.1.6
139
A Faulty Induction Proof
If we have done a good job in writing this text, right about now you should be
thinking, “Hey, this induction stuff isn’t so hard after all—just show P .0/ is true
and that P .n/ implies P .n C 1/ for any number n.” And, you would be right,
although sometimes when you start doing induction proofs on your own, you can
run into trouble. For example, we will now use induction to “prove” that all horses
are the same color—just when you thought it was safe to skip class and work on
your robot program instead. Sorry!
False Theorem. All horses are the same color.
Notice that no n is mentioned in this assertion, so we’re going to have to reformulate it in a way that makes an n explicit. In particular, we’ll (falsely) prove
that
False Theorem 5.1.3. In every set of n 1 horses, all the horses are the same
color.
This is a statement about all integers n 1 rather 0, so it’s natural to use a
slight variation on induction: prove P .1/ in the base case and then prove that P .n/
implies P .n C 1/ for all n 1 in the inductive step. This is a perfectly valid variant
of induction and is not the problem with the proof below.
Bogus proof. The proof is by induction on n. The induction hypothesis P .n/ will
be
In every set of n horses, all are the same color.
(5.3)
Base case: (n D 1). P .1/ is true, because in a size-1 set of horses, there’s only one
horse, and this horse is definitely the same color as itself.
Inductive step: Assume that P .n/ is true for some n 1. That is, assume that in
every set of n horses, all are the same color. Now suppose we have a set of n C 1
horses:
h1 ; h2 ; : : : ; hn ; hnC1 :
We need to prove these n C 1 horses are all the same color.
By our assumption, the first n horses are the same color:
h ; h ; : : : ; hn ; hnC1
„1 2ƒ‚
…
same color
Also by our assumption, the last n horses are the same color:
h1 ; h2 ; : : : ; hn ; hnC1
„
ƒ‚
…
same color
“mcs” — 2017/6/5 — 19:42 — page 140 — #148
140
Chapter 5
Induction
So h1 is the same color as the remaining horses besides hnC1 —that is, h2 ; : : : ; hn .
Likewise, hnC1 is the same color as the remaining horses besides h1 —that is,
h2 ; : : : ; hn , again. Since h1 and hnC1 are the same color as h2 ; : : : ; hn , all n C 1
horses must be the same color, and so P .n C 1/ is true. Thus, P .n/ implies
P .n C 1/.
By the principle of induction, P .n/ is true for all n 1.
We’ve proved something false! Does this mean that math broken and we should
all take up poetry instead? Of course not! It just means that this proof has a mistake.
The mistake in this argument is in the sentence that begins “So h1 is the same
color as the remaining horses besides hnC1 —that is h2 ; : : : ; hn ; : : : .” The ellipis
notation (“: : : ”) in the expression “h1 ; h2 ; : : : ; hn ; hnC1 ” creates the impression
that there are some remaining horses—namely h2 ; : : : ; hn —besides h1 and hnC1 .
However, this is not true when n D 1. In that case, h1 ; h2 ; : : : ; hn ; hnC1 is just
h1 ; h2 and there are no “remaining” horses for h1 to share a color with. And of
course, in this case h1 and h2 really don’t need to be the same color.
This mistake knocks a critical link out of our induction argument. We proved
P .1/ and we correctly proved P .2/ ! P .3/, P .3/ ! P .4/, etc. But we failed
to prove P .1/ ! P .2/, and so everything falls apart: we cannot conclude that
P .2/, P .3/, etc., are true. And naturally, these propositions are all false; there are
sets of n horses of different colors for all n 2.
Students sometimes explain that the mistake in the proof is because P .n/ is false
for n 2, and the proof assumes something false, P .n/, in order to prove P .nC1/.
You should think about how to help such a student understand why this explanation
would get no credit on a Math for Computer Science exam.
5.2
Strong Induction
A useful variant of induction is called strong induction. Strong induction and ordinary induction are used for exactly the same thing: proving that a predicate is true
for all nonnegative integers. Strong induction is useful when a simple proof that
the predicate holds for n C 1 does not follow just from the fact that it holds at n,
but from the fact that it holds for other values n.
“mcs” — 2017/6/5 — 19:42 — page 141 — #149
5.2. Strong Induction
5.2.1
141
A Rule for Strong Induction
Principle of Strong Induction.
Let P be a predicate on nonnegative integers. If
P .0/ is true, and
for all n 2 N, P .0/, P .1/, . . . , P .n/ together imply P .n C 1/,
then P .m/ is true for all m 2 N.
The only change from the ordinary induction principle is that strong induction
allows you make more assumptions in the inductive step of your proof! In an
ordinary induction argument, you assume that P .n/ is true and try to prove that
P .n C 1/ is also true. In a strong induction argument, you may assume that P .0/,
P .1/, . . . , and P .n/ are all true when you go to prove P .nC1/. So you can assume
a stronger set of hypotheses which can make your job easier.
Formulated as a proof rule, strong induction is
Rule. Strong Induction Rule
P .0/;
8n 2 N: P .0/ AND P .1/ AND : : : AND P .n/ IMPLIES P .n C 1/
8m 2 N: P .m/
Stated more succintly, the rule is
Rule.
P .0/;
Œ8k n 2 N: P .k/ IMPLIES P .n C 1/
8m 2 N: P .m/
The template for strong induction proofs is identical to the template given in
Section 5.1.3 for ordinary induction except for two things:
you should state that your proof is by strong induction, and
you can assume that P .0/, P .1/, . . . , P .n/ are all true instead of only P .n/
during the inductive step.
5.2.2
Fibonacci numbers
The numbers that bear his name arose out of the Italian mathematician Fibonacci’s
models of population growth at the beginning of the thirteenth century. Fibonacci
numbers turn out to describe the growth of lots of interesting biological quantities
“mcs” — 2017/6/5 — 19:42 — page 142 — #150
142
Chapter 5
Induction
such as the shape of pineapple sprouts or pine cones, and they also come up regularly in Computer Science where they describe the growth of various data structures
and computation times of algorithms.
To generate the list of successive Fibonacci numbers, you start by writing 0; 1
and then keep adding another element to the list by summing the two previous
ones:
0; 1; 1; 2; 3; 5; 8; 13; 21; : : : :
Another way to describe this process is to define nth Fibonacci number F .n/ by
the equations:
F .0/ WWD 0;
F .1/ WWD 1;
F .n/ WWD F .n
1/ C F .n
2/
for n 2.
Note that because the general rule for finding the Fibonacci F .n/ refers to the two
previous values F .n 1/ and F .n 2/, we needed to know the two values F .0/
and F .1/ in order to get started.
One simple property of Fibonacci numbers is that the even/odd pattern of Fibonacci numbers repeats in a cycle of length three. A nice way to say this is that
for all n 0,
F .n/ is even IFF F .n C 3/ is even:
(5.4)
We will verify the equivalence (5.4) by induction, but strong induction is called for
because properties of F .n/ depend not just on F .n 1/ but also on F .n 2/.
Proof. The (strong) induction hypothesis P .n/ will be (5.4).
Base cases:
(n D 0). F .0/ D 0 and F .3/ D 2 are both even.
(n D 1). F .1/ D 1 and F .4/ D 3 are both not even.
Induction step: For n 1, we want to prove P .n C 1/ is true assuming by strong
induction that P .n/ and P .n 1/ are true.
Now it is easy to verify that for all integers k; m,
m C k is even IFF Œm is even IFF k is even:
(*)
“mcs” — 2017/6/5 — 19:42 — page 143 — #151
5.2. Strong Induction
143
So for n 1,
F .n C 1/ is even
IFF
F .n/ C F .n
(def of F .n C 1/)
IFF
ŒF .n/ is even IFF F .n
IFF
ŒF .n C 3/ is even IFF F .n C 2/ is even
1/ is even
1/ is even
(by strong ind. hyp. P .n/; P .n
IFF
F .n C 3/ C F .n C 2/ is even
IFF
F .n C 4/ is even
(by (*))
1/)
(by (*))
(by def of F .n C 4/):
This shows that
F .n C 1/ is even IFF F .n C 4/ is even;
which means that P .n C 1/ is true, as required.
There is a long standing community of Fibonacci number enthusiasts who have
been captivated by the many extraordinary properties of these number—a few further illustrative properties appear in Problems 5.8, 5.25, and 5.30.
5.2.3
Products of Primes
We can use strong induction to re-prove Theorem 2.3.1 which we previously proved
using Well Ordering.
Theorem. Every integer greater than 1 is a product of primes.
Proof. We will prove the Theorem by strong induction, letting the induction hypothesis P .n/ be
n is a product of primes:
So the Theorem will follow if we prove that P .n/ holds for all n 2.
Base Case: (n D 2): P .2/ is true because 2 is prime, so it is a length one product
of primes by convention.
Inductive step: Suppose that n 2 and that every number from 2 to n is a product
of primes. We must show that P .n C 1/ holds, namely, that n C 1 is also a product
of primes. We argue by cases:
If n C 1 is itself prime, then it is a length one product of primes by convention,
and so P .n C 1/ holds in this case.
Otherwise, n C 1 is not prime, which by definition means n C 1 D k m for some
integers k; m between 2 and n. Now by the strong induction hypothesis, we know
that both k and m are products of primes. By multiplying these products, it follows
“mcs” — 2017/6/5 — 19:42 — page 144 — #152
144
Chapter 5
Induction
Figure 5.5 One way to make 26 Sg using Strongian currency
immediately that k m D n C 1 is also a product of primes. Therefore, P .n C 1/
holds in this case as well.
So P .n C 1/ holds in any case, which completes the proof by strong induction
that P .n/ holds for all n 2.
5.2.4
Making Change
The country Inductia, whose unit of currency is the Strong, has coins worth 3Sg
(3 Strongs) and 5Sg. Although the Inductians have some trouble making small
change like 4Sg or 7Sg, it turns out that they can collect coins to make change for
any number that is at least 8 Strongs.
Strong induction makes this easy to prove for n C 1 11, because then .n C
1/ 3 8, so by strong induction the Inductians can make change for exactly
.n C 1/ 3 Strongs, and then they can add a 3Sg coin to get .n C 1/Sg. So the only
thing to do is check that they can make change for all the amounts from 8 to 10Sg,
which is not too hard to do.
Here’s a detailed writeup using the official format:
Proof. We prove by strong induction that the Inductians can make change for any
amount of at least 8Sg. The induction hypothesis P .n/ will be:
There is a collection of coins whose value is n C 8 Strongs.
We now proceed with the induction proof:
Base case: P .0/ is true because a 3Sg coin together with a 5Sg coin makes 8Sg.
“mcs” — 2017/6/5 — 19:42 — page 145 — #153
5.2. Strong Induction
145
Inductive step: We assume P .k/ holds for all k n, and prove that P .n C 1/
holds. We argue by cases:
Case (n C 1 = 1): We have to make .n C 1/ C 8 D 9Sg. We can do this using
three 3Sg coins.
Case (n C 1 = 2): We have to make .n C 1/ C 8 D 10Sg. Use two 5Sg coins.
Case (n C 1 3): Then 0 n 2 n, so by the strong induction hypothesis,
the Inductians can make change for .n 2/ C 8Sg. Now by adding a 3Sg coin, they
can make change for .n C 1/ C 8Sg, so P .n C 1/ holds in this case.
Since n 0, we know that n C 1 1 and thus that the three cases cover
every possibility. Since P .n C 1/ is true in every case, we can conclude by strong
induction that for all n 0, the Inductians can make change for n C 8 Strong. That
is, they can make change for any number of eight or more Strong.
5.2.5
The Stacking Game
Here is another exciting game that’s surely about to sweep the nation!
You begin with a stack of n boxes. Then you make a sequence of moves. In each
move, you divide one stack of boxes into two nonempty stacks. The game ends
when you have n stacks, each containing a single box. You earn points for each
move; in particular, if you divide one stack of height a C b into two stacks with
heights a and b, then you score ab points for that move. Your overall score is the
sum of the points that you earn for each move. What strategy should you use to
maximize your total score?
As an example, suppose that we begin with a stack of n D 10 boxes. Then the
game might proceed as shown in Figure 5.6. Can you find a better strategy?
Analyzing the Game
Let’s use strong induction to analyze the unstacking game. We’ll prove that your
score is determined entirely by the number of boxes—your strategy is irrelevant!
Theorem 5.2.1. Every way of unstacking n blocks gives a score of n.n
points.
1/=2
There are a couple technical points to notice in the proof:
The template for a strong induction proof mirrors the one for ordinary induction.
As with ordinary induction, we have some freedom to adjust indices. In this
case, we prove P .1/ in the base case and prove that P .1/; : : : ; P .n/ imply
P .n C 1/ for all n 1 in the inductive step.
“mcs” — 2017/6/5 — 19:42 — page 146 — #154
146
Chapter 5
Induction
Stack Heights
10
5
5
4
2
2
1
1
1
1
5
3
3
3
2
2
1
1
1
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
2
2
2
2
2
1
1
1
1
1
1
Score
25 points
6
4
4
2
1
1
1 1
1
1 1 1
1
1 1 1 1
1
Total Score D 45 points
Figure 5.6 An example of the stacking game with n D 10 boxes. On each line,
the underlined stack is divided in the next step.
Proof. The proof is by strong induction. Let P .n/ be the proposition that every
way of unstacking n blocks gives a score of n.n 1/=2.
Base case: If n D 1, then there is only one block. No moves are possible, and so
the total score for the game is 1.1 1/=2 D 0. Therefore, P .1/ is true.
Inductive step: Now we must show that P .1/, . . . , P .n/ imply P .n C 1/ for all
n 1. So assume that P .1/, . . . , P .n/ are all true and that we have a stack of
n C 1 blocks. The first move must split this stack into substacks with positive sizes
a and b where a C b D n C 1 and 0 < a; b n. Now the total score for the game
is the sum of points for this first move plus points obtained by unstacking the two
resulting substacks:
total score D (score for 1st move)
C (score for unstacking a blocks)
C (score for unstacking b blocks)
a.a 1/ b.b 1/
D ab C
C
2
2
.a C b/2 .a C b/
.a C b/..a C b/
D
D
2
2
.n C 1/n
D
2
This shows that P .1/, P .2/, . . . , P .n/ imply P .n C 1/.
by P .a/ and P .b/
1/
“mcs” — 2017/6/5 — 19:42 — page 147 — #155
5.3. Strong Induction vs. Induction vs. Well Ordering
Therefore, the claim is true by strong induction.
5.3
147
Strong Induction vs. Induction vs. Well Ordering
Strong induction looks genuinely “stronger” than ordinary induction —after all,
you can assume a lot more when proving the induction step. Since ordinary induction is a special case of strong induction, you might wonder why anyone would
bother with the ordinary induction.
But strong induction really isn’t any stronger, because a simple text manipulation program can automatically reformat any proof using strong induction into a
proof using ordinary induction—just by decorating the induction hypothesis with
a universal quantifier in a standard way. Still, it’s worth distinguishing these two
kinds of induction, since which you use will signal whether the inductive step for
n C 1 follows directly from the case for n or requires cases smaller than n, and that
is generally good for your reader to know.
The template for the two kinds of induction rules looks nothing like the one for
the Well Ordering Principle, but this chapter included a couple of examples where
induction was used to prove something already proved using well ordering. In fact,
this can always be done. As the examples may suggest, any well ordering proof
can automatically be reformatted into an induction proof. So theoretically, no one
need bother with the Well Ordering Principle either.
But it’s equally easy to go the other way, and automatically reformat any strong
induction proof into a Well Ordering proof. The three proof methods—well ordering, induction, and strong induction—are simply different formats for presenting
the same mathematical reasoning!
So why three methods? Well, sometimes induction proofs are clearer because
they don’t require proof by contradiction. Also, induction proofs often provide
recursive procedures that reduce large inputs to smaller ones. On the other hand,
well ordering can come out slightly shorter and sometimes seem more natural and
less worrisome to beginners.
So which method should you use? There is no simple recipe. Sometimes the
only way to decide is to write up a proof using more than one method and compare
how they come out. But whichever method you choose, be sure to state the method
up front to help a reader follow your proof.
“mcs” — 2017/6/5 — 19:42 — page 148 — #156
148
Chapter 5
Induction
Figure 5.7 Gehry’s new tile.
Problems for Section 5.1
Practice Problems
Problem 5.1.
Prove by induction that every nonempty finite set of real numbers has a minimum
element.
Problem 5.2.
Frank Gehry has changed his mind. Instead of the L-shaped tiles shown in figure 5.3, he wants to use an odd offset pattern of tiles (or its mirror-image reflection),
as shown in 5.7. To prove this is possible, he uses reasoning similar to the proof
in 5.1.5. However, unlike the proof in the text, this proof is flawed. Which part of
the proof below contains a logical error?
False Claim. The proof is by induction. Let P .n/ be the proposition that for every
location of Bill in a 2n 2n courtyard, there exists a tiling of the remainder with
the offset tile pattern.
False proof. Base case: P .0/ is true because Bill fills the whole courtyard.
Inductive step: Assume that P .n/ is true for some n 0; that is, for every location
of Bill in a 2n 2n courtyard, there exists a tiling of the remainder. Divide the
2nC1 2nC1 courtyard into four quadrants, each 2n 2n . One quadrant contains
“mcs” — 2017/6/5 — 19:42 — page 149 — #157
5.3. Strong Induction vs. Induction vs. Well Ordering
149
Figure 5.8 The induction hypothesis for the false theorem.
Bill (B in the diagram below). Place a temporary Bill (X in the diagram) in each of
the three squares lying near this quadrant as shown in Figure 5.8.
We can tile each of the four quadrants by the induction assumption. Replacing
the three temporary Bills with a single offset tile completes the job. This proves
that P .n/ implies P .n C 1/ for all n 0. Thus P .m/ is true for all m 2 N, and the
ability to place Bill in the center of the courtyard follows as a special case where
we put Bill in a central square.
Class Problems
Problem 5.3.
Use induction to prove that
13 C 23 C C n3 D
for all n 1.
Remember to formally
1. Declare proof by induction.
2. Identify the induction hypothesis P .n/.
3. Establish the base case.
n.n C 1/
2
2
:
(5.5)
“mcs” — 2017/6/5 — 19:42 — page 150 — #158
150
Chapter 5
Induction
4. Prove that P .n/ ) P .n C 1/.
5. Conclude that P .n/ holds for all n 1.
as in the five part template.
Problem 5.4.
Prove by induction on n that
1 C r C r2 C C rn D
r nC1 1
r 1
(5.6)
for all n 2 N and numbers r ¤ 1.
Problem 5.5.
Prove by induction:
1C
1
1 1
C C C 2 < 2
4 9
n
1
;
n
(5.7)
for all n > 1.
Problem 5.6. (a) Prove by induction that a 2n 2n courtyard with a 1 1 statue
of Bill in a corner can be covered with L-shaped tiles. (Do not assume or reprove
the (stronger) result of Theorem 5.1.2 that Bill can be placed anywhere. The point
of this problem is to show a different induction hypothesis that works.)
(b) Use the result of part (a) to prove the original claim that there is a tiling with
Bill in the middle.
Problem 5.7.
We’ve proved in two different ways that
n.n C 1/
2
But now we’re going to prove a contradictory theorem!
1 C 2 C 3 C C n D
False Theorem. For all n 0,
2 C 3 C 4 C C n D
n.n C 1/
2
“mcs” — 2017/6/5 — 19:42 — page 151 — #159
5.3. Strong Induction vs. Induction vs. Well Ordering
151
Proof. We use induction. Let P .n/ be the proposition that 2 C 3 C 4 C C n D
n.n C 1/=2.
Base case: P .0/ is true, since both sides of the equation are equal to zero. (Recall
that a sum with no terms is zero.)
Inductive step: Now we must show that P .n/ implies P .n C 1/ for all n 0. So
suppose that P .n/ is true; that is, 2 C 3 C 4 C C n D n.n C 1/=2. Then we can
reason as follows:
2 C 3 C 4 C C n C .n C 1/ D Œ2 C 3 C 4 C C n C .n C 1/
n.n C 1/
D
C .n C 1/
2
.n C 1/.n C 2/
D
2
Above, we group some terms, use the assumption P .n/, and then simplify. This
shows that P .n/ implies P .n C 1/. By the principle of induction, P .n/ is true for
all n 2 N.
Where exactly is the error in this proof?
Homework Problems
Problem 5.8.
The Fibonacci numbers F .n/ are described in Section 5.2.2.
Prove by induction that for all n 1,
F .n
1/ F .n C 1/
F .n/2 D . 1/n :
(5.8)
Problem 5.9.
For any binary string ˛ let num .˛/ be the nonnegative integer it represents in binary
notation. For example, num .10/ D 2, and num .0101/ D 5.
An n C 1-bit adder adds two n C 1-bit binary numbers. More precisely, an
n C 1-bit adder takes two length n C 1 binary strings
˛n WWD an : : : a1 a0 ;
ˇn WWD bn : : : b1 b0 ;
and a binary digit c0 as inputs, and produces a length-(n C 1) binary string
n WWD sn : : : s1 s0 ;
“mcs” — 2017/6/5 — 19:42 — page 152 — #160
152
Chapter 5
Induction
and a binary digit cnC1 as outputs, and satisfies the specification:
num .˛n / C num .ˇn / C c0 D 2nC1 cnC1 C num .n / :
(5.9)
There is a straighforward way to implement an nC1-bit adder as a digital circuit:
an n C 1-bit ripple-carry circuit has 1 C 2.n C 1/ binary inputs
an ; : : : ; a1 ; a0 ; bn ; : : : ; b1 ; b0 ; c0 ;
and n C 2 binary outputs,
cnC1 ; sn ; : : : ; s1 ; s0 :
As in Problem 3.6, the ripple-carry circuit is specified by the following formulas:
si WWD ai XOR bi XOR ci
ci C1 WWD .ai AND bi / OR .ai AND ci / OR .bi AND ci /; :
(5.10)
(5.11)
for 0 i n.
(a) Verify that definitions (5.10) and (5.11) imply that
an C bn C cn D 2cnC1 C sn :
(5.12)
for all n 2 N.
(b) Prove by induction on n that an n C 1-bit ripple-carry circuit really is an n C 1bit adder, that is, its outputs satisfy (5.9).
Hint: You may assume that, by definition of binary representation of integers,
num .˛nC1 / D anC1 2nC1 C num .˛n / :
(5.13)
Problem 5.10.
Divided Equilateral Triangles5 (DETs) can be built up as follows:
A single equilateral triangle counts as a DET whose only subtriangle is itself.
is a DET, then the equilateral triangle T 0 built out of four
If T WWD
copies of T as shown in in Figure 5.9 is also a DET, and the subtriangles of
T 0 are exactly the subtriangles of each of the copies of T .
5 Adapted
from [49].
“mcs” — 2017/6/5 — 19:42 — page 153 — #161
5.3. Strong Induction vs. Induction vs. Well Ordering
153
[h]
Figure 5.9 DET T 0 from Four Copies of DET T
[h]
Figure 5.10 Trapezoid from Three Triangles
(a) Define the length of a DET to be the number of subtriangles with an edge on
its base. Prove by induction on length that the total number of subtriangles of a
DET is the square of its length.
(b) Show that a DET with one of its corner subtriangles removed can be tiled with
trapezoids built out of three subtriangles as in Figure 5.10.
Problem 5.11.
The Math for Computer Science mascot, Theory Hippotamus, made a startling
discovery while playing with his prized collection of unit squares over the weekend.
Here is what happened.
First, Theory Hippotamus put his favorite unit square down on the floor as in
Figure 5.11 (a). He noted that the length of the periphery of the resulting shape was
4, an even number. Next, he put a second unit square down next to the first so that
the two squares shared an edge as in Figure 5.11 (b). He noticed that the length
of the periphery of the resulting shape was now 6, which is also an even number.
(The periphery of each shape in the figure is indicated by a thicker line.) Theory
Hippotamus continued to place squares so that each new square shared an edge with
at least one previously-placed square and no squares overlapped. Eventually, he
arrived at the shape in Figure 5.11 (c). He realized that the length of the periphery
of this shape was 36, which is again an even number.
Our plucky porcine pal is perplexed by this peculiar pattern. Use induction on
the number of squares to prove that the length of the periphery is always even, no
matter how many squares Theory Hippotamus places or how he arranges them.
“mcs” — 2017/6/5 — 19:42 — page 154 — #162
154
Chapter 5
Induction
(a)
(b)
(c)
Figure 5.11 Some shapes that Theory Hippotamus created.
Problem 5.12.
Prove the Distributive Law of intersection over the union of n sets by induction:
A\
n
[
i D1
Bi D
n
[
.A \ Bi /:
(5.14)
i D1
Hint: Theorem 4.1.2 gives the n D 2 case.
Problem 5.13.
Here is an interesting construction of a geometric object known as the Koch snowflake.
Define a sequence of polygons S0 ; S1 recursively, starting with S0 equal to an equilateral triangle with unit sides. We construct SnC1 by removing the middle third
of each edge of Sn and replacing it with two line segments of the same length, as
illustrated in Figure 5.12.
Let an be the area of Sn . Observe that p
a0 is just the area of the unit equilateral
triangle which by elementary geometry is 3=4.
Prove by induction that for n 0, the area of the nth snowflake is given by:
8 3 4 n
an D a0
:
(5.15)
5 5 9
“mcs” — 2017/6/5 — 19:42 — page 155 — #163
5.3. Strong Induction vs. Induction vs. Well Ordering
155
Figure 5.12 S0 ; S1 ; S2 and S3 .
Exam Problems
Problem 5.14.
Prove by induction that
n
X
k kŠ D .n C 1/Š
1:
(5.16)
1
Problem 5.15.
Prove by induction:
03 C 13 C 23 C C n3 D
n.n C 1/
2
2
using the equation itself as the induction hypothesis P .n/.
(a) Prove the
base case .n D 0/.
(b) Now prove the
inductive step.
; 8n 0:
“mcs” — 2017/6/5 — 19:42 — page 156 — #164
156
Chapter 5
Induction
Problem 5.16.
Suppose P .n/ is a predicate on nonnegative numbers, and suppose
8k: P .k/ IMPLIES P .k C 2/:
(5.17)
For P ’s that satisfy (5.17), some of the assertions below Can hold for some,
but not all, such P , other assertions Always hold no matter what the P may be,
and some Never hold for any such P . Indicate which case applies for each of the
assertions and briefly explain why.
(a) 8n 0: P .n/
(b) NOT.P .0// AND 8n 1: P .n/
(c) 8n 0: NOT.P .n//
(d) .8n 100: P .n// AND .8n > 100: NOT.P .n///
(e) .8n 100: NOT.P .n/// AND .8n > 100: P .n//
(f) P .0/ IMPLIES 8n: P .n C 2/
(g) Œ9n: P .2n/ IMPLIES 8n: P .2n C 2/
(h) P .1/ IMPLIES 8n: P .2n C 1/
(i) Œ9n: P .2n/ IMPLIES 8n: P .2n C 2/
(j) 9n: 9m > n: ŒP .2n/ AND NOT.P .2m//
(k) Œ9n: P .n/ IMPLIES 8n: 9m > n: P .m/
(l) NOT.P .0// IMPLIES 8n: NOT.P .2n//
Problem 5.17.
We examine a series of propositional formulas F1 ; F2 ; : : : ; Fn ; : : : containing propositional variables P1 ; P2 ; : : : ; Pn ; : : : constructed as follows
F1 .P1 /
F2 .P1 ; P2 /
F3 .P1 ; P2 ; P3 /
F4 .P1 ; P2 ; P3 ; P4 /
F5 .P1 ; P2 ; P3 ; P4 ; P5 /
WWD
WWD
WWD
WWD
WWD
::
:
P1
P1 IMPLIES P2
.P1 IMPLIES P2 / IMPLIES P3
..P1 IMPLIES P2 / IMPLIES P3 / IMPLIES P4
...P1 IMPLIES P2 / IMPLIES P3 / IMPLIES P4 / IMPLIES P5
“mcs” — 2017/6/5 — 19:42 — page 157 — #165
5.3. Strong Induction vs. Induction vs. Well Ordering
157
Let Tn be the number of different true/false settings of the variables P1 ; P2 ; : : : ; Pn
for which Fn .P1 ; P2 ; : : : ; Pn / is true. For example, T2 D 3 since F2 .P1 ; P2 / is
true for 3 different settings of the variables P1 and P2 :
P1 P2 F2 .P1 ; P2 /
T T
T
T F
F
F T
T
F F
T
(a) Explain why
TnC1 D 2nC1
Tn :
(5.18)
2nC1 C . 1/n
3
(*)
(b) Use induction to prove that
Tn D
for n 1.
Problem 5.18.
You are given n envelopes, numbered 0; 1; : : : ; n 1. Envelope 0 contains 20 D 1
dollar, Envelope 1 contains 21 D 2 dollars, . . . , and Envelope n 1 contains 2n 1
dollars. Let P .n/ be the assertion that:
For all nonnegative integers k < 2n , there is a subset of the n envelopes
whose contents total to exactly k dollars.
Prove by induction that P .n/ holds for all integers n 1.
Problem 5.19.
Prove by induction that
1 2 C 2 3 C 3 4 C C n.n C 1/ D
for all integers n 1.
n.n C 1/.n C 2/
3
(5.19)
“mcs” — 2017/6/5 — 19:42 — page 158 — #166
158
Chapter 5
Induction
...
AND-circuit
NOT-gate
Figure 5.13
OR -circuit
from AND-circuit.
Problem 5.20.
A k-bit AND-circuit is a digital circuit that has k 0-1 valued inputs6 d0 ; d1 ; : : : ; dk
and one 0-1-valued output variable whose value will be
d0 AND d1 AND AND dk
1
1:
OR -circuits
are defined in the same way, with “OR” replacing “AND.”
(a) Suppose we want an OR-circuit but only have a supply of AND-circuits and
some NOT-gates (“inverters”) that have one 0-1 valued input and one 0-1 valued
output. We can turn an AND-circuit into an OR-circuit by attaching a NOT-gate to
each input of the AND-circuit and also attaching a NOT-gate to the output of the
AND -circuit. This is illustrated in Figure 5.13. Briefly explain why this works.
Large digital circuits are built by connecting together smaller digital circuits as
components. One of the most basic components is a two-input/one-output ANDgate that produces an output value equal to the AND of its two input values. So
according the definition in part (a), a single AND-gate is a 1-bit AND-circuit.
We can build up larger AND-circuits out of a collection of AND-gates in several
ways. For example, one way to build a 4-bit AND-circuit is to connect three ANDgates as illustrated in Figure 5.14.
More generally, a depth-n tree-design AND-circuit—“depth-n circuit” for short—
has 2n inputs and is built from two depth-.n 1/ circuits by using the outputs of
the two depth-.n 1/ circuits as inputs to a single AND-gate. This is illustrated in
Figure 5.15. So the 4-bit AND-circuit in Figure 5.14 is a depth-2 circuit. A depth-1
circuit is defined simply to be a single AND-gate.
6 Following
F.
the usual conventions for digital circuits, we’re using 1 for the truth value T and 0 for
“mcs” — 2017/6/5 — 19:42 — page 159 — #167
5.3. Strong Induction vs. Induction vs. Well Ordering
a2
AND
159
a3
AND
AND
Figure 5.14 A 4-bit AND-circuit.
a2n-1-1
depthn-1
circuit
a2n-1 a2n-1+1 •.. a2n-1
depthn-1
circuit
AND
Figure 5.15 An n-bit tree-design AND-circuit.
“mcs” — 2017/6/5 — 19:42 — page 160 — #168
160
Chapter 5
Induction
(b) Let gate#.n/ be the number of AND-gates in a depth-n circuit. Prove by induction that
gate#.n/ D 2n 1
(5.20)
for all n 1.
Problems for Section 5.2
Practice Problems
Problem 5.21.
Some fundamental principles for reasoning about nonnegative integers are:
1. The Induction Principle,
2. The Strong Induction Principle,
3. The Well Ordering Principle.
Identify which, if any, of the above principles is captured by each of the following
inference rules.
(a)
P .0/; 8m: .8k m: P .k// IMPLIES P .m C 1/
8n: P .n/
(b)
P .b/; 8k b: P .k/ IMPLIES P .k C 1/
8k b: P .k/
(c)
9n: P .n/
9m: ŒP .m/ AND .8k: P .k/ IMPLIES k m/
(d)
P .0/; 8k > 0: P .k/ IMPLIES P .k C 1/
8n: P .n/
(e)
8m: .8k < m: P .k// IMPLIES P .m/
8n: P .n/
“mcs” — 2017/6/5 — 19:42 — page 161 — #169
5.3. Strong Induction vs. Induction vs. Well Ordering
161
Problem 5.22.
The Fibonacci numbers F .n/ are described in Section 5.2.2.
Indicate exactly which sentence(s) in the following bogus proof contain logical
errors? Explain.
False Claim. Every Fibonacci number is even.
Bogus proof. Let all the variables n; m; k mentioned below be nonnegative integer
valued. Let Even.n/ mean that F .n/ is even. The proof is by strong induction with
induction hypothesis Even.n/.
base case: F .0/ D 0 is an even number, so Even.0/ is true.
inductive step: We assume may assume the strong induction hypothesis
Even.k/ for 0 k n;
and we must prove Even.n C 1/.
Then by strong induction hypothesis, Even.n/ and Even.n 1/ are true, that
is, F .n/ and F .n 1/ are both even. But by the definition, F .n C 1/ equals the
sum F .n/ C F .n 1/ of two even numbers, and so it is also even. This proves
Even.n C 1/ as required.
Hence, F .m/ is even for all m 2 N by the Strong Induction Principle.
Problem 5.23.
Alice wants to prove by induction that a predicate P holds for certain nonnegative
integers. She has proven that for all nonnegative integers n D 0; 1; : : :
P .n/ IMPLIES P .n C 3/:
(a) Suppose Alice also proves that P .5/ holds. Which of the following propositions can she infer?
1. P .n/ holds for all n 5
2. P .3n/ holds for all n 5
3. P .n/ holds for n D 8; 11; 14; : : :
4. P .n/ does not hold for n < 5
5. 8n: P .3n C 5/
6. 8n > 2: P .3n
1/
“mcs” — 2017/6/5 — 19:42 — page 162 — #170
162
Chapter 5
Induction
7. P .0/ IMPLIES 8n: P .3n C 2/
8. P .0/ IMPLIES 8n: P .3n/
(b) Which of the following could Alice prove in order to conclude that P .n/ holds
for all n 5?
1. P .0/
2. P .5/
3. P .5/ and P .6/
4. P .0/, P .1/ and P .2/
5. P .5/, P .6/ and P .7/
6. P .2/, P .4/ and P .5/
7. P .2/, P .4/ and P .6/
8. P .3/, P .5/ and P .7/
Problem 5.24.
Prove that every amount of postage of 12 cents or more can be formed using just
4-cent and 5-cent stamps.
Class Problems
Problem 5.25.
The Fibonacci numbers are described in Section 5.2.2.
Prove, using strong induction, the following closed-form formula for the Fibonacci numbers.7
pn qn
F .n/ D p
5
p
p
where p D 1C2 5 and q D 1 2 5 .
Hint: Note that p and q are the roots of x 2
2
q D q C 1.
x
1 D 0, and so p 2 D p C 1 and
Problem 5.26.
A sequence of numbers is weakly decreasing when each number in the sequence is
7 This mind-boggling formula is known as Binet’s formula.
in Chapter 22, how it comes about.
We’ll explain in Chapter 16, and again
“mcs” — 2017/6/5 — 19:42 — page 163 — #171
5.3. Strong Induction vs. Induction vs. Well Ordering
163
the numbers after it. (This implies that a sequence of just one number is weakly
decreasing.)
Here’s a bogus proof of a very important true fact, every integer greater than 1 is
a product of a unique weakly decreasing sequence of primes—a pusp, for short.
Explain what’s bogus about the proof.
Lemma. Every integer greater than 1 is a pusp.
For example, 252 D 7 3 3 2 2, and no other weakly decreasing sequence of
primes will have a product equal to 252.
Bogus proof. We will prove the lemma by strong induction, letting the induction
hypothesis P .n/ be
n is a pusp:
So the lemma will follow if we prove that P .n/ holds for all n 2.
Base Case (n D 2): P .2/ is true because 2 is prime, and so it is a length one
product of primes, and this is obviously the only sequence of primes whose product
can equal 2.
Inductive step: Suppose that n 2 and that i is a pusp for every integer i where
2 i < n C 1. We must show that P .n C 1/ holds, namely, that n C 1 is also a
pusp. We argue by cases:
If n C 1 is itself prime, then it is the product of a length one sequence consisting
of itself. This sequence is unique, since by definition of prime, n C 1 has no other
prime factors. So n C 1 is a pusp, that is P .n C 1/ holds in this case.
Otherwise, n C 1 is not prime, which by definition means n C 1 D km for
some integers k; m such that 2 k; m < n C 1. Now by the strong induction
hypothesis, we know that k and m are pusps. It follows that by merging the unique
prime sequences for k and m, in sorted order, we get a unique weakly decreasing
sequence of primes whose product equals n C 1. So n C 1 is a pusp, in this case as
well.
So P .n C 1/ holds in any case, which completes the proof by strong induction
that P .n/ holds for all n 2.
Problem 5.27.
Define the potential p.S / of a stack of blocks S to be k.k 1/=2 where k is the
number of blocks in S. Define the potential p.A/ of a set of stacks A to be the sum
of the potentials of the stacks in A.
“mcs” — 2017/6/5 — 19:42 — page 164 — #172
164
Chapter 5
Induction
Generalize Theorem 5.2.1 about scores in the stacking game to show that for any
set of stacks A if a sequence of moves starting with A leads to another set of stacks
B then p.A/ p.B/, and the score for this sequence of moves is p.A/ p.B/.
Hint: Try induction on the number of moves to get from A to B.
Homework Problems
Problem 5.28.
A group of n 1 people can be divided into teams, each containing either 4 or
7 people. What are all the possible values of n? Use induction to prove that your
answer is correct.
Problem 5.29.
The following Lemma is true, but the proof given for it below is defective. Pinpoint exactly where the proof first makes an unjustified step and explain why it is
unjustified.
Lemma. For any prime p and positive integers n; x1 ; x2 ; : : : ; xn , if p j x1 x2 : : : xn ,
then p j xi for some 1 i n.
Bogus proof. Proof by strong induction on n. The induction hypothesis P .n/ is
that Lemma holds for n.
Base case n D 1: When n D 1, we have p j x1 , therefore we can let i D 1 and
conclude p j xi .
Induction step: Now assuming the claim holds for all k n, we must prove it
for n C 1.
So suppose p j x1 x2 xnC1 . Let yn D xn xnC1 , so x1 x2 xnC1 D x1 x2 xn
Since the right-hand side of this equality is a product of n terms, we have by induction that p divides one of them. If p j xi for some i < n, then we have the desired
i . Otherwise p j yn . But since yn is a product of the two terms xn ; xnC1 , we have
by strong induction that p divides one of them. So in this case p j xi for i D n or
i D n C 1.
Exam Problems
Problem 5.30.
The Fibonacci numbers F .n/ are described in Section 5.2.2.
These numbers satisfy many unexpected identities, such as
F .0/2 C F .1/2 C C F .n/2 D F .n/F .n C 1/:
(5.21)
1 yn .
“mcs” — 2017/6/5 — 19:42 — page 165 — #173
5.3. Strong Induction vs. Induction vs. Well Ordering
165
Equation (5.21) can be proved to hold for all n 2 N by induction, using the equation
itself as the induction hypothesis P .n/.
(a) Prove the
base case .n D 0/.
(b) Now prove the
inductive step.
Problem 5.31.
Use strong induction to prove that n 3n=3 for every integer n 0.
Problem 5.32.
A class of any size of 18 or more can be assembled from student teams of sizes 4
and 7. Prove this by induction (of some kind), using the induction hypothesis:
S.n/ WWD a class of n C 18 students can be assembled from teams of sizes 4 and 7:
Problem 5.33.
Any amount of ten or more cents postage that is a multiple of five can be made
using only 10¢ and 15¢ stamps. Prove this by induction (ordinary or strong, but say
which) using the induction hypothesis
S.n/ WWD .5n C 10/¢ postage can be made using only 10¢ and 15¢ stamps:
“mcs” — 2017/6/5 — 19:42 — page 166 — #174
“mcs” — 2017/6/5 — 19:42 — page 167 — #175
6
State Machines
State machines are a simple, abstract model of step-by-step processes. Since computer programs can be understood as defining step-by-step computational processes,
it’s not surprising that state machines come up regularly in computer science. They
also come up in many other settings such as designing digital circuits and modeling probabilistic processes. This section introduces Floyd’s Invariant Principle
which is a version of induction tailored specifically for proving properties of state
machines.
One of the most important uses of induction in computer science involves proving one or more desirable properties continues to hold at every step in a process.
A property that is preserved through a series of operations or steps is known as a
preserved invariant.
Examples of desirable invariants include properties such as a variable never exceeding a certain value, the altitude of a plane never dropping below 1,000 feet
without the wingflaps being deployed, and the temperature of a nuclear reactor
never exceeding the threshold for a meltdown.
6.1
States and Transitions
Formally, a state machine is nothing more than a binary relation on a set, except
that the elements of the set are called “states,” the relation is called the transition
relation, and an arrow in the graph of the transition relation is called a transition.
A transition from state q to state r will be written q ! r. The transition relation
is also called the state graph of the machine. A state machine also comes equipped
with a designated start state.
A simple example is a bounded counter, which counts from 0 to 99 and overflows
at 100. This state machine is pictured in Figure 6.1, with states pictured as circles,
transitions by arrows, and with start state 0 indicated by the double circle. To be
start
state
0
1
2
99
overflow
Figure 6.1 State transitions for the 99-bounded counter.
“mcs” — 2017/6/5 — 19:42 — page 168 — #176
168
Chapter 6
State Machines
precise, what the picture tells us is that this bounded counter machine has
states WWD f0; 1; : : : ; 99; overflowg;
start state WWD 0;
transitions WWD fn ! n C 1 j 0 n < 99g
[ f99 ! overflow; overflow ! overflowg:
This machine isn’t much use once it overflows, since it has no way to get out of its
overflow state.
State machines for digital circuits and string pattern matching algorithms, for instance, usually have only a finite number of states. Machines that model continuing
computations typically have an infinite number of states. For example, instead of
the 99-bounded counter, we could easily define an “unbounded” counter that just
keeps counting up without overflowing. The unbounded counter has an infinite
state set, the nonnegative integers, which makes its state diagram harder to draw.
State machines are often defined with labels on states and/or transitions to indicate such things as input or output values, costs, capacities, or probabilities. Our
state machines don’t include any such labels because they aren’t needed for our
purposes. We do name states, as in Figure 6.1, so we can talk about them, but the
names aren’t part of the state machine.
6.2
The Invariant Principle
6.2.1
A Diagonally-Moving Robot
Suppose we have a robot that starts at the origin and moves on an infinite 2dimensional integer grid. The state of the robot at any time can be specified by
the integer coordinates .x; y/ of the robot’s current position. So the start state
is .0; 0/. At each step, the robot may move to a diagonally adjacent grid point, as
illustrated in Figure 6.2.
To be precise, the robot’s transitions are:
f.m; n/ ! .m ˙ 1; n ˙ 1/ j m; n 2 Zg:
For example, after the first step, the robot could be in states .1; 1/, .1; 1/, . 1; 1/
or . 1; 1/. After two steps, there are 9 possible states for the robot, including .0; 0/. The question is, can the robot ever reach position .1; 0/?
If you play around with the robot a bit, you’ll probably notice that the robot can
only reach positions .m; n/ for which m C n is even, which of course means that it
“mcs” — 2017/6/5 — 19:42 — page 169 — #177
6.2. The Invariant Principle
169
y
2
1
0
x
0
1
2
3
Figure 6.2 The Diagonally Moving Robot.
can’t reach .1; 0/. This follows because the evenness of the sum of the coordinates
is a property that is preserved by transitions. This is an example of a preserved
invariant.
This once, let’s go through this preserved invariant argument, carefully highlighting where induction comes in. Specifically, define the even-sum property of
states to be:
Even-sum..m; n// WWD Œm C n is even:
Lemma 6.2.1. For any transition q ! r of the diagonally-moving robot, if Evensum(q), then Even-sum(r).
This lemma follows immediately from the definition of the robot’s transitions:
.m; n/ ! .m ˙ 1; n ˙ 1/. After a transition, the sum of coordinates changes by
.˙1/ C .˙1/, that is, by 0, 2, or -2. Of course, adding 0, 2 or -2 to an even number
gives an even number. So by a trivial induction on the number of transitions, we
can prove:
Theorem 6.2.2. The sum of the coordinates of any state reachable by the diagonallymoving robot is even.
“mcs” — 2017/6/5 — 19:42 — page 170 — #178
170
Chapter 6
State Machines
y
‹‹
2
1
0
goal
x
0
1
2
3
Figure 6.3 Can the Robot get to .1; 0/?
“mcs” — 2017/6/5 — 19:42 — page 171 — #179
6.2. The Invariant Principle
171
Proof. The proof is induction on the number of transitions the robot has made. The
induction hypothesis is
P .n/ WWD if q is a state reachable in n transitions, then Even-sum(q):
Base case: P .0/ is true since the only state reachable in 0 transitions is the start
state .0; 0/, and 0 C 0 is even.
Inductive step: Assume that P .n/ is true, and let r be any state reachable in n C 1
transitions. We need to prove that Even-sum(r) holds.
Since r is reachable in n C 1 transitions, there must be a state q reachable in n
transitions such that q ! r. Since P .n/ is assumed to be true, Even-sum(q) holds,
and so by Lemma 6.2.1, Even-sum(r) also holds. This proves that P .n/ IMPLIES
P .n C 1/ as required, completing the proof of the inductive step.
We conclude by induction that for all n 0, if q is reachable in n transitions, then
Even-sum(q). This implies that every reachable state has the Even-sum property.
Corollary 6.2.3. The robot can never reach position .1; 0/.
Proof. By Theorem 6.2.2, we know the robot can only reach positions with coordinates that sum to an even number, and thus it cannot reach position .1; 0/.
6.2.2
Statement of the Invariant Principle
Using the Even-sum invariant to understand the diagonally-moving robot is a simple example of a basic proof method called The Invariant Principle. The Principle
summarizes how induction on the number of steps to reach a state applies to invariants.
A state machine execution describes a possible sequence of steps a machine
might take.
Definition 6.2.4. An execution of the state machine is a (possibly infinite) sequence
of states with the property that
it begins with the start state, and
if q and r are consecutive states in the sequence, then q ! r.
A state is called reachable if it appears in some execution.
Definition 6.2.5. A preserved invariant of a state machine is a predicate P on
states, such that whenever P .q/ is true of a state q and q ! r for some state r
then P .r/ holds.
“mcs” — 2017/6/5 — 19:42 — page 172 — #180
172
Chapter 6
State Machines
The Invariant Principle
If a preserved invariant of a state machine is true for the start state,
then it is true for all reachable states.
The Invariant Principle is nothing more than the Induction Principle reformulated
in a convenient form for state machines. Showing that a predicate is true in the start
state is the base case of the induction, and showing that a predicate is a preserved
invariant corresponds to the inductive step.1
1 Preserved
invariants are commonly just called “invariants” in the literature on program correctness, but we decided to throw in the extra adjective to avoid confusion with other definitions. For
example, other texts (as well as another subject at MIT) use “invariant” to mean “predicate true of
all reachable states.” Let’s call this definition “invariant-2.” Now invariant-2 seems like a reasonable definition, since unreachable states by definition don’t matter, and all we want to show is that
a desired property is invariant-2. But this confuses the objective of demonstrating that a property is
invariant-2 with the method of finding a preserved invariant—which is preserved even at unreachable
states—to show that it is invariant-2.
“mcs” — 2017/6/5 — 19:42 — page 173 — #181
6.2. The Invariant Principle
173
Robert W. Floyd
The Invariant Principle was formulated by Robert W. Floyd at Carnegie Tech
in 1967. (Carnegie Tech was renamed Carnegie-Mellon University the following
year.) Floyd was already famous for work on the formal grammars that transformed the field of programming language parsing; that was how he got to be
a professor even though he never got a Ph.D. (He had been admitted to a PhD
program as a teenage prodigy, but flunked out and never went back.)
In that same year, Albert R. Meyer was appointed Assistant Professor in the
Carnegie Tech Computer Science Department, where he first met Floyd. Floyd
and Meyer were the only theoreticians in the department, and they were both delighted to talk about their shared interests. After just a few conversations, Floyd’s
new junior colleague decided that Floyd was the smartest person he had ever met.
Naturally, one of the first things Floyd wanted to tell Meyer about was his new,
as yet unpublished, Invariant Principle. Floyd explained the result to Meyer, and
Meyer wondered (privately) how someone as brilliant as Floyd could be excited
by such a trivial observation. Floyd had to show Meyer a bunch of examples before Meyer understood Floyd’s excitement —not at the truth of the utterly obvious
Invariant Principle, but rather at the insight that such a simple method could be so
widely and easily applied in verifying programs.
Floyd left for Stanford the following year. He won the Turing award—the
“Nobel prize” of computer science—in the late 1970’s, in recognition of his work
on grammars and on the foundations of program verification. He remained at
Stanford from 1968 until his death in September, 2001. You can learn more about
Floyd’s life and work by reading the eulogy at
http://oldwww.acm.org/pubs/membernet/stories/floyd.pdf
written by his closest colleague, Don Knuth.
“mcs” — 2017/6/5 — 19:42 — page 174 — #182
174
Chapter 6
6.2.3
State Machines
The Die Hard Example
The movie Die Hard 3: With a Vengeance includes an amusing example of a state
machine. The lead characters played by Samuel L. Jackson and Bruce Willis have
to disarm a bomb planted by the diabolical Simon Gruber:
Simon: On the fountain, there should be 2 jugs, do you see them? A 5gallon and a 3-gallon. Fill one of the jugs with exactly 4 gallons of water
and place it on the scale and the timer will stop. You must be precise;
one ounce more or less will result in detonation. If you’re still alive in 5
minutes, we’ll speak.
Bruce: Wait, wait a second. I don’t get it. Do you get it?
Samuel: No.
Bruce: Get the jugs. Obviously, we can’t fill the 3-gallon jug with 4 gallons of water.
Samuel: Obviously.
Bruce: All right. I know, here we go. We fill the 3-gallon jug exactly to
the top, right?
Samuel: Uh-huh.
Bruce: Okay, now we pour this 3 gallons into the 5-gallon jug, giving us
exactly 3 gallons in the 5-gallon jug, right?
Samuel: Right, then what?
Bruce: All right. We take the 3-gallon jug and fill it a third of the way...
Samuel: No! He said, “Be precise.” Exactly 4 gallons.
Bruce: Sh - -. Every cop within 50 miles is running his a - - off and I’m
out here playing kids games in the park.
Samuel: Hey, you want to focus on the problem at hand?
Fortunately, they find a solution in the nick of time. You can work out how.
The Die Hard 3 State Machine
The jug-filling scenario can be modeled with a state machine that keeps track of
the amount b of water in the big jug, and the amount l in the little jug. With the 3
and 5 gallon water jugs, the states formally will be pairs .b; l/ of real numbers such
“mcs” — 2017/6/5 — 19:42 — page 175 — #183
6.2. The Invariant Principle
175
that 0 b 5; 0 l 3. (We can prove that the reachable values of b and l will
be nonnegative integers, but we won’t assume this.) The start state is .0; 0/, since
both jugs start empty.
Since the amount of water in the jug must be known exactly, we will only consider moves in which a jug gets completely filled or completely emptied. There are
several kinds of transitions:
1. Fill the little jug: .b; l/ ! .b; 3/ for l < 3.
2. Fill the big jug: .b; l/ ! .5; l/ for b < 5.
3. Empty the little jug: .b; l/ ! .b; 0/ for l > 0.
4. Empty the big jug: .b; l/ ! .0; l/ for b > 0.
5. Pour from the little jug into the big jug: for l > 0,
(
.b C l; 0/
if b C l 5,
.b; l/ !
.5; l .5 b// otherwise.
6. Pour from big jug into little jug: for b > 0,
(
.0; b C l/
if b C l 3,
.b; l/ !
.b .3 l/; 3/ otherwise.
Note that in contrast to the 99-counter state machine, there is more than one possible transition out of states in the Die Hard machine. Machines like the 99-counter
with at most one transition out of each state are called deterministic. The Die Hard
machine is nondeterministic because some states have transitions to several different states.
The Die Hard 3 bomb gets disarmed successfully because the state (4,3) is reachable.
Die Hard Permanently
The Die Hard series is getting tired, so we propose a final Die Hard Permanently.
Here, Simon’s brother returns to avenge him, posing the same challenge, but with
the 5 gallon jug replaced by a 9 gallon one. The state machine has the same specification as the Die Hard 3 version, except all occurrences of “5” are replaced by
“9.”
Now, reaching any state of the form .4; l/ is impossible. We prove this using
the Invariant Principle. Specifically, we define the preserved invariant predicate
P ..b; l// to be that b and l are nonnegative integer multiples of 3.
“mcs” — 2017/6/5 — 19:42 — page 176 — #184
176
Chapter 6
State Machines
To prove that P is a preserved invariant of Die-Hard-Once-and-For-All machine,
we assume P .q/ holds for some state q WWD .b; l/ and that q ! r. We have to
show that P .r/ holds. The proof divides into cases, according to which transition
rule is used.
One case is a “fill the little jug” transition. This means r D .b; 3/. But P .q/
implies that b is an integer multiple of 3, and of course 3 is an integer multiple of
3, so P .r/ still holds.
Another case is a “pour from big jug into little jug” transition. For the subcase
when there isn’t enough room in the little jug to hold all the water, that is, when
b C l > 3, we have r D .b .3 l/; 3/. But P .q/ implies that b and l are integer
multiples of 3, which means b .3 l/ is too, so in this case too, P .r/ holds.
We won’t bother to crank out the remaining cases, which can all be checked
just as easily. Now by the Invariant Principle, we conclude that every reachable
state satisifies P . But since no state of the form .4; l/ satisifies P , we have proved
rigorously that Bruce dies once and for all!
By the way, notice that the state (1,0), which satisfies NOT.P /, has a transition
to (0,0), which satisfies P . So the negation of a preserved invariant may not be a
preserved invariant.
6.3
Partial Correctness & Termination
Floyd distinguished two required properties to verify a program. The first property
is called partial correctness; this is the property that the final results, if any, of the
process must satisfy system requirements.
You might suppose that if a result was only partially correct, then it might also
be partially incorrect, but that’s not what Floyd meant. The word “partial” comes
from viewing a process that might not terminate as computing a partial relation.
Partial correctness means that when there is a result, it is correct, but the process
might not always produce a result, perhaps because it gets stuck in a loop.
The second correctness property, called termination, is that the process does
always produce some final value.
Partial correctness can commonly be proved using the Invariant Principle. Termination can commonly be proved using the Well Ordering Principle. We’ll illustrate
this by verifying a Fast Exponentiation procedure.
“mcs” — 2017/6/5 — 19:42 — page 177 — #185
6.3. Partial Correctness & Termination
6.3.1
177
Fast Exponentiation
Exponentiating
The most straightforward way to compute the bth power of a number a is to multiply a by itself b 1 times. But the solution can be found in considerably fewer multiplications by using a technique called Fast Exponentiation. The register machine
program below defines the fast exponentiation algorithm. The letters x; y; z; r denote registers that hold numbers. An assignment statement has the form “z WD a”
and has the effect of setting the number in register z to be the number a.
A Fast Exponentiation Program
Given inputs a 2 R; b 2 N, initialize registers x; y; z to a; 1; b respectively, and
repeat the following sequence of steps until termination:
if z D 0 return y and terminate
r WD remainder.z; 2/
z WD quotient.z; 2/
if r D 1, then y WD xy
x WD x 2
We claim this program always terminates and leaves y D ab .
To begin, we’ll model the behavior of the program with a state machine:
1. states WWD R R N,
2. start state WWD .a; 1; b/,
3. transitions are defined by the rule
(
.x 2 ; y; quotient.z; 2//
if z is nonzero and even;
.x; y; z/ !
2
.x ; xy; quotient.z; 2// if z is nonzero and odd:
The preserved invariant P ..x; y; z// will be
z 2 N AND yx z D ab :
(6.1)
To prove that P is preserved, assume P ..x; y; z// holds and that .x; y; z/ !
.xt ; yt ; zt /. We must prove that P ..xt ; yt ; zt // holds, that is,
zt 2 N AND yt xtzt D ab :
(6.2)
“mcs” — 2017/6/5 — 19:42 — page 178 — #186
178
Chapter 6
State Machines
Since there is a transition from .x; y; z/, we have z ¤ 0, and since z 2 N
by (6.1), we can consider just two cases:
If z is even, then we have that xt D x 2 ; yt D y; zt D z=2. Therefore, zt 2 N
and
yt xtzt D y.x 2 /z=2
D yx 2z=2
D yx z
D ab
(by (6.1))
If z is odd, then we have that xt D x 2 ; yt D xy; zt D .z
zt 2 N and
yt xtzt D xy.x 2 /.z
1/=2
D yx 1C2.z
1/=2
D yx 1C.z
1/=2. Therefore,
1/
D yx z
D ab
(by (6.1))
So in both cases, (6.2) holds, proving that P is a preserved invariant.
Now it’s easy to prove partial correctness: if the Fast Exponentiation program
terminates, it does so with ab in register y. This works because 1 ab D ab , which
means that the start state .a; 1; b/ satisifies P . By the Invariant Principle, P holds
for all reachable states. But the program only stops when z D 0. If a terminated
state .x; y; 0/ is reachable, then y D yx 0 D ab as required.
Ok, it’s partially correct, but what’s fast about it? The answer is that the number
of multiplications it performs to compute ab is roughly the length of the binary
representation of b. That is, the Fast Exponentiation program uses roughly log b 2
multiplications, compared to the naive approach of multiplying by a a total of b 1
times.
More precisely, it requires at most 2.dlog be C 1/ multiplications for the Fast
Exponentiation algorithm to compute ab for b > 1. The reason is that the number
in register z is initially b, and gets at least halved with each transition. So it can’t
be halved more than dlog be C 1 times before hitting zero and causing the program
to terminate. Since each of the transitions involves at most two multiplications, the
total number of multiplications until z D 0 is at most 2.dlog be C 1/ for b > 0 (see
Problem 6.6).
2 As
usual in computer science, log b means the base two logarithm log2 b. We use, ln b for the
natural logarithm loge b, and otherwise write the logarithm base explicitly, as in log10 b.
“mcs” — 2017/6/5 — 19:42 — page 179 — #187
6.3. Partial Correctness & Termination
6.3.2
179
Derived Variables
The preceding termination proof involved finding a nonnegative integer-valued
measure to assign to states. We might call this measure the “size” of the state.
We then showed that the size of a state decreased with every state transition. By
the Well Ordering Principle, the size can’t decrease indefinitely, so when a minimum size state is reached, there can’t be any transitions possible: the process has
terminated.
More generally, the technique of assigning values to states—not necessarily nonnegative integers and not necessarily decreasing under transitions—is often useful
in the analysis of algorithms. Potential functions play a similar role in physics. In
the context of computational processes, such value assignments for states are called
derived variables.
For example, for the Die Hard machines we could have introduced a derived
variable f W states ! R for the amount of water in both buckets, by setting
f ..a; b// WWD a C b. Similarly, in the robot problem, the position of the robot along
the x-axis would be given by the derived variable x-coord, where x-coord..i; j //WWD i .
There are a few standard properties of derived variables that are handy in analyzing state machines.
Definition 6.3.1. A derived variable f W states ! R is strictly decreasing iff
q ! q 0 IMPLIES f .q 0 / < f .q/:
It is weakly decreasing iff
q ! q 0 IMPLIES f .q 0 / f .q/:
Strictly increasingweakly increasing derived variables are defined similarly.3
We confirmed termination of the Fast Exponentiation procedure by noticing that
the derived variable z was nonnegative-integer-valued and strictly decreasing. We
can summarize this approach to proving termination as follows:
Theorem 6.3.2. If f is a strictly decreasing N-valued derived variable of a state
machine, then the length of any execution starting at state q is at most f .q/.
Of course, we could prove Theorem 6.3.2 by induction on the value of f .q/, but
think about what it says: “If you start counting down at some nonnegative integer
f .q/, then you can’t count down more than f .q/ times.” Put this way, it’s obvious.
3 Weakly increasing variables are often also called nondecreasing. We will avoid this terminology
to prevent confusion between nondecreasing variables and variables with the much weaker property
of not being a decreasing variable.
“mcs” — 2017/6/5 — 19:42 — page 180 — #188
180
Chapter 6
6.3.3
State Machines
Termination with Well ordered Sets (Optional)
Theorem 6.3.2 generalizes straightforwardly to derived variables taking values in a
well ordered set (Section 2.4.
Theorem 6.3.3. If there exists a strictly decreasing derived variable whose range
is a well ordered set, then every execution terminates.
Theorem 6.3.3 follows immediately from the observation that a set of numbers
is well ordered iff it has no infinite decreasing sequences (Problem 2.23).
Note that the existence of a weakly decreasing derived variable does not guarantee that every execution terminates. An infinite execution could proceed through
states in which a weakly decreasing variable remained constant.
6.3.4
A Southeast Jumping Robot (Optional)
Here’s a simple, contrived example of a termination proof based on a variable that
is strictly decreasing over a well ordered set. Let’s think about a robot that travels
around the nonnegative integer quadrant N2 .
If the robot is at some position .x; y/ different from the origin .0; 0/, the robot
must make a move, which may be
a unit distance West—that is, .x; y/ ! .x
1; y/ for x > 0, or
a unit distance South combined with an arbitrary jump East—that is, .x; y/ !
.z; y 1/ for z x,
providing the move does not leave the quadrant.
Claim 6.3.4. The robot will always get stuck at the origin.
If we think of the robot as a nondeterministic state machine, then Claim 6.3.4 is
a termination assertion. The Claim may seem obvious, but it really has a different
character than termination based on nonnegative integer-valued variables. That’s
because, even knowing that the robot is at position .0; 1/, for example, there is no
way to bound the time it takes for the robot to get stuck. It can delay getting stuck
for as many seconds as it wants by making its next move to a distant point in the
Far East. This rules out proving termination using Theorem 6.3.2.
So does Claim 6.3.4 still seem obvious?
Well it is if you see the trick. Define a derived variable v mapping robot states
to the numbers in the well ordered set N C F of Lemma 2.4.6. In particular, define
v W N2 ! N C F as follows
x
v.x; y/ WWD y C
:
xC1
“mcs” — 2017/6/5 — 19:42 — page 181 — #189
6.4. The Stable Marriage Problem
181
2
Brad
1
Jennifer
1
2
2
1
Billy Bob
1
2
Angelina
Figure 6.4 Preferences for four people. Both men like Angelina best and both
women like Brad best.
Now it’s easy to check that if .x; y/ ! .x 0 ; y 0 / is a legitimate robot move, then
v..x 0 ; y 0 // < v..x; y//. In particular, v is a strictly decreasing derived variable, so
Theorem 6.3.3 implies that the robot always get stuck—even though we can’t say
how many moves it will take until it does.
6.4
The Stable Marriage Problem
Suppose we have a population of men and women in which each person has preferences of the opposite-gender person they would like to marry: each man has his
preference list of all the women, and each woman has her preference list of all of
the men.
The preferences don’t have to be symmetric. That is, Jennifer might like Brad
best, but Brad doesn’t necessarily like Jennifer best. The goal is to marry everyone: every man must marry exactly one woman and vice versa—no polygamy and
heterosexual marriages only.4 Moreover, we would like to find a matching between
men and women that is stable in the sense that there is no pair of people who prefer
one another to their spouses.
For example, suppose Brad likes Angelina best, and Angelina likes Brad best, but
Brad and Angelina are married to other people, say Jennifer and Billy Bob. Now
Brad and Angelina prefer each other to their spouses, which puts their marriages
at risk. Pretty soon, they’re likely to start spending late nights together working on
problem sets!
This unfortunate situation is illustrated in Figure 6.4, where the digits “1” and “2”
near a man shows which of the two women he ranks first and second, respectively,
and similarly for the women.
4 Same-sex
marriage is an interesting but separate case.
“mcs” — 2017/6/5 — 19:42 — page 182 — #190
182
Chapter 6
State Machines
More generally, in any matching, a man and woman who are not married to each
other and who like each other better than their spouses is called a rogue couple. In
the situation shown in Figure 6.4, Brad and Angelina would be a rogue couple.
Having a rogue couple is not a good thing, since it threatens the stability of the
marriages. On the other hand, if there are no rogue couples, then for any man and
woman who are not married to each other, at least one likes their spouse better than
the other, and so there won’t be any mutual temptation to start an affair.
Definition 6.4.1. A stable matching is a matching with no rogue couples.
The question is, given everybody’s preferences, can you find a stable set of marriages? In the example consisting solely of the four people in Figure 6.4, we could
let Brad and Angelina both have their first choices by marrying each other. Now
neither Brad nor Angelina prefers anybody else to their spouse, so neither will be
in a rogue couple. This leaves Jen not-so-happily married to Billy Bob, but neither
Jen nor Billy Bob can entice somebody else to marry them, and so this is a stable
matching.
It turns out there always is a stable matching among a group of men and women.
We don’t know of any immediate way to recognize this, and it seems surprising. In
fact, in the apparently similar same-sex or “buddy” matching problem where people
are supposed to be paired off as buddies, regardless of gender, a stable matching
may not be possible. An example of preferences among four people where there is
no stable buddy match is given in Problem 6.22. But when men are only allowed
to marry women, and vice versa, then there is a simple procedure to produce a
stable matching and the concept of preserved invariants provides an elegant way to
understand and verify the procedure.
6.4.1
The Mating Ritual
The procedure for finding a stable matching can be described in a memorable way
as a Mating Ritual that takes place over several days. On the starting day, each man
has his full preference list of all the women, and likewise each woman has her full
preference list of all the men. Then following events happen each day:
Morning: Each man stands under the balcony of the woman on the top of his
list, that is the woman he prefers above all the other remaining women. Then he
serenades her. He is said to be her suitor. If a man has no women left on his list, he
stays home and does his math homework.
Afternoon: Each woman who has one or more suitors says to her favorite among
them, “We might get engaged. Please stay around.” To the other suitors, she says,
“No. I will never marry you! Take a hike!”
“mcs” — 2017/6/5 — 19:42 — page 183 — #191
6.4. The Stable Marriage Problem
183
Evening: Any man who is told by a woman to take a hike crosses that woman
off his preference list.
Termination condition: When a day arrives in which every woman has at most
one suitor, the ritual ends with each woman marrying her suitor, if she has one.
There are a number of facts about this Mating Ritual that we would like to prove:
The Ritual eventually reaches the termination condition.
Everybody ends up married.
The resulting marriages are stable.
To prove these facts, it will be helpful to recognize the Ritual as the description
of a state machine. The state at the start of any day is determined by knowing for
each man, which woman, if any, he will serenade that day—that is, the woman
at the top of his preference list after he has crossed out all the women who have
rejected him on earlier days.
Mating Ritual at Akamai
The Internet infrastructure company Akamai, cofounded by Tom Leighton, also
uses a variation of the Mating Ritual to assign web traffic to its servers.
In the early days, Akamai used other combinatorial optimization algorithms
that got to be too slow as the number of servers (over 65,000 in 2010) and requests
(over 800 billion per day) increased. Akamai switched to a Ritual-like approach,
since a Ritual is fast and can be run in a distributed manner. In this case, web
requests correspond to women and web servers correspond to men. The web
requests have preferences based on latency and packet loss, and the web servers
have preferences based on cost of bandwidth and co-location.
“mcs” — 2017/6/5 — 19:42 — page 184 — #192
184
Chapter 6
6.4.2
State Machines
There is a Marriage Day
It’s easy to see why the Mating Ritual has a terminal day when people finally get
married. Every day on which the ritual hasn’t terminated, at least one man crosses
a woman off his list. (If the ritual hasn’t terminated, there must be some woman
serenaded by at least two men, and at least one of them will have to cross her off his
list). If we start with n men and n women, then each of the n men’s lists initially
has n women on it, for a total of n2 list entries. Since no women ever gets added
to a list, the total number of entries on the lists decreases every day that the Ritual
continues, and so the Ritual can continue for at most n2 days.
6.4.3
They All Live Happily Ever After. . .
We will prove that the Mating Ritual leaves everyone in a stable marriage. To do
this, we note one very useful fact about the Ritual: if on some morning a woman has
any suitor, then her favorite suitor will still be serenading her the next morning—his
list won’t have changed. So she is sure to have today’s favorite suitor among her
suitors tomorrow. That means she will be able to choose a favorite suitor tomorrow
who is at least as desirable to her as today’s favorite. So day by day, her favorite
suitor can stay the same or get better, never worse. This sounds like an invariant,
and it is. Namely, let P be the predicate
For every woman w and man m, if w is crossed off m’s list, then w has
a suitor whom she prefers over m.
Lemma 6.4.2. P is a preserved invariant for The Mating Ritual.
Proof. Woman w gets crossed off m’s list only when w has a suitor she prefers to
m. Thereafter, her favorite suitor doesn’t change until one she likes better comes
along. So if her favorite suitor was preferable to m, then any new favorite suitor
will be as well.
Notice that the invariant P holds vacuously at the beginning since no women are
crossed off to start. So by the Invariant Principle, P holds throughout the Ritual.
Now we can prove:
Theorem 6.4.3. Everyone is married at the end of the Mating Ritual.
Proof. Assume to the contrary that on the last day of the Mating Ritual, some
man—call him Bob—is not married. This means Bob can’t be serenading anybody,
that is, his list must be empty. So every woman must have been crossed off his
list and, since P is true, every woman has a suitor whom she prefers to Bob. In
“mcs” — 2017/6/5 — 19:42 — page 185 — #193
6.4. The Stable Marriage Problem
185
particular, every woman has some suitor, and since it is the last day, they have only
one suitor, and this is who they marry. But there are an equal number of men and
women, so if all women are married, so are all men, contradicting the assumption
that Bob is not married.
Theorem 6.4.4. The Mating Ritual produces a stable matching.
Proof. Let Brad and Jen be any man and woman, respectively, that are not married
to each other on the last day of the Mating Ritual. We will prove that Brad and Jen
are not a rogue couple, and thus that all marriages on the last day are stable. There
are two cases to consider.
Case 1: Jen is not on Brad’s list by the end. Then by invariant P , we know that
Jen has a suitor (and hence a husband) whom she prefers to Brad. So she’s
not going to run off with Brad—Brad and Jen cannot be a rogue couple.
Case 2: Jen is on Brad’s list. Since Brad picks women to serenade by working
down his list, his wife must be higher on his preference list than Jen. So
he’s not going to run off with Jen—once again, Brad and Jen are not a rogue
couple.
6.4.4
. . . Especially the Men
Who is favored by the Mating Ritual, the men or the women? The women seem
to have all the power: each day they choose their favorite suitor and reject the rest.
What’s more, we know their suitors can only change for the better as the Ritual
progresses. Similarly, a man keeps serenading the woman he most prefers among
those on his list until he must cross her off, at which point he serenades the next
most preferred woman on his list. So from the man’s perspective, the woman he is
serenading can only change for the worse. Sounds like a good deal for the women.
But it’s not! We will show that the men are by far the favored gender under the
Mating Ritual.
While the Mating Ritual produces one stable matching, stable matchings need
not be unique. For example, reversing the roles of men and women will often yield
a different stable matching among them. So a man may have different wives in
different sets of stable marriages. In some cases, a man can stably marry every one
of the women, but in most cases, there are some women who cannot be a man’s wife
in any stable matching. For example, given the preferences shown in Figure 6.4,
Jennifer cannot be Brad’s wife in any stable matching because if he was married to
her, then he and Angelina would be a rogue couple. It is not feasible for Jennifer to
be stably married to Brad.
“mcs” — 2017/6/5 — 19:42 — page 186 — #194
186
Chapter 6
State Machines
Definition 6.4.5. Given a set of preferences for the men and women, one person is
a feasible spouse for another person when there is a stable matching in which these
two people are married.
Definition 6.4.6. Let Q be the predicate: for every woman w and man m, if w is
crossed off m’s list, then w is not a feasible spouse for m.
Lemma 6.4.7. Q is a preserved invariant5 for The Mating Ritual.
Proof. Suppose Q holds at some point in the Ritual and some woman Alice is
about to be crossed off some man’s, Bob’s, list. We claim that Alice must not be
feasible for Bob. Therefore Q will still hold after Alice is crossed off, proving that
Q is invariant.
To verify the claim, notice that when Alice gets crossed of Bob’s list, it’s because
Alice has a suitor, Ted, she prefers to Bob. What’s more, since Q holds, all Ted’s
feasible wives are still on his list, and Alice is at the top. So Ted likes Alice better
than all his other feasible spouses. Now if Alice could be married to Bob in some
set of stable marriages, then Ted must be married to a wife he likes less than Alice,
making Alice and Ted a rogue couple and contradicting stability. So Alice can’t be
married to Bob, that is, Alice is not a feasible wife for Bob, as claimed.
Definition 6.4.8. Given a set of preferences for the men and women, a person’s
optimal spouse is their most preferred feasible spouse. A person’s pessimal spouse
is their least preferred feasible spouse.
Everybody has an optimal and a pessimal spouse, since we know there is at least
one stable matching, namely, the one produced by the Mating Ritual. Lemma 6.4.7
implies a key property the Mating Ritual:
Theorem 6.4.9. The Mating Ritual marries every man to his optimal wife.
Proof. If Bob is married to Alice on the final day of the Ritual, then everyone above
Alice on Bob’s preference list was crossed off, and by property Q, all these crossed
off women were infeasible for Bob. So Alice is Bob’s highest ranked feasible
spouse, that is, his optimal wife.
Lemma 6.4.10. Given a set of preferences for the men and women, everyone is the
pessimal spouse of their optimal spouse.
5 We appeal to P in justifying Q, so technically it is P
invariant. But let’s not be picky.
AND
Q which is actually the preserved
“mcs” — 2017/6/5 — 19:42 — page 187 — #195
6.4. The Stable Marriage Problem
187
Proof. By symmetry, it is enough to prove that every man is the pessimal husband
of his optimal wife.
Suppose Alice is Bob’s optimal wife. Then in any stable set of marriages, if Alice
liked her husband less Bob, then Bob must be married to someone not optimal, and
therefore Alice and Bob would be a rogue couple. So Alice must like all her feasible
husbands at least as much as Bob. That is, Bob is Alice’s pessimal husband.
Corollary 6.4.11. The Mating Ritual marries every woman to her pessimal husband.
6.4.5
Applications
The Mating Ritual was first announced in a paper by D. Gale and L.S. Shapley in
1962, but ten years before the Gale-Shapley paper was published, and unknown to
them, a similar algorithm was being used to assign residents to hospitals by the National Resident Matching Program (NRMP). The NRMP has, since the turn of the
twentieth century, assigned each year’s pool of medical school graduates to hospital residencies (formerly called “internships”), with hospitals and graduates playing
the roles of men and women.6 Before the Ritual-like algorithm was adopted, there
were chronic disruptions and awkward countermeasures taken to preserve unstable assignments of graduates to residencies. The Ritual resolved these problems so
successfully, that it was used essentially without change at least through 1989.7 For
this and related work, Shapley was awarded the 2012 Nobel prize in Economics.
Not surprisingly, the Mating Ritual is also used by at least one large online dating agency. Of course there is no serenading going on—everything is handled by
computer.
Problems for Section 6.3
Practice Problems
Problem 6.1.
Which states of the Die Hard 3 machine below have transitions to exactly two
states?
6 In this case there may be multiple women married to one man, but this is a minor complication,
see Problem 6.23.
7 Much more about the Stable Marriage Problem can be found in the very readable mathematical
monograph by Dan Gusfield and Robert W. Irving, [27].
“mcs” — 2017/6/5 — 19:42 — page 188 — #196
188
Chapter 6
State Machines
Die Hard Transitions
1. Fill the little jug: .b; l/ ! .b; 3/ for l < 3.
2. Fill the big jug: .b; l/ ! .5; l/ for b < 5.
3. Empty the little jug: .b; l/ ! .b; 0/ for l > 0.
4. Empty the big jug: .b; l/ ! .0; l/ for b > 0.
5. Pour from the little jug into the big jug: for l > 0,
(
.b C l; 0/
if b C l 5,
.b; l/ !
.5; l .5 b// otherwise.
6. Pour from big jug into little jug: for b > 0,
(
.0; b C l/
if b C l 3,
.b; l/ !
.b .3 l/; 3/ otherwise.
Homework Problems
Problem 6.2.
In the late 1960s, the military junta that ousted the government of the small republic of Nerdia completely outlawed built-in multiplication operations, and also
forbade division by any number other than 3. Fortunately, a young dissident found
a way to help the population multiply any two nonnegative integers without risking
persecution by the junta. The procedure he taught people is:
procedure multiply.x; y: nonnegative integers/
r WD x;
s WD y;
a WD 0;
while s ¤ 0 do
if 3 j s then
r WD r C r C r;
s WD s=3;
else if 3 j .s 1/ then
a WD a C r;
r WD r C r C r;
s WD .s 1/=3;
else
“mcs” — 2017/6/5 — 19:42 — page 189 — #197
6.4. The Stable Marriage Problem
189
a WD a C r C r;
r WD r C r C r;
s WD .s 2/=3;
return a;
We can model the algorithm as a state machine whose states are triples of nonnegative integers .r; s; a/. The initial state is .x; y; 0/. The transitions are given by
the rule that for s > 0:
8
ˆ
if 3 j s
<.3r; s=3; a/
.r; s; a/ ! .3r; .s 1/=3; a C r/
if 3 j .s 1/
ˆ
:
.3r; .s 2/=3; a C 2r/ otherwise:
(a) List the sequence of steps that appears in the execution of the algorithm for
inputs x D 5 and y D 10.
(b) Use the Invariant Method to prove that the algorithm is partially correct—that
is, if s D 0, then a D xy.
(c) Prove that the algorithm terminates after at most 1 C log3 y executions of the
body of the do statement.
Problem 6.3.
A robot named Wall-E wanders around a two-dimensional grid. He starts out at
.0; 0/ and is allowed to take four different types of steps:
1. .C2; 1/
2. .C1; 2/
3. .C1; C1/
4. . 3; 0/
Thus, for example, Wall-E might walk as follows. The types of his steps are
listed above the arrows.
1
3
2
4
.0; 0/ ! .2; 1/ ! .3; 0/ ! .4; 2/ ! .1; 2/ ! : : :
Wall-E’s true love, the fashionable and high-powered robot, Eve, awaits at .0; 2/.
“mcs” — 2017/6/5 — 19:42 — page 190 — #198
190
Chapter 6
State Machines
(a) Describe a state machine model of this problem.
(b) Will Wall-E ever find his true love? Either find a path from Wall-E to Eve, or
use the Invariant Principle to prove that no such path exists.
Problem 6.4.
A hungry ant is placed on an unbounded grid. Each square of the grid either contains a crumb or is empty. The squares containing crumbs form a path in which,
except at the ends, every crumb is adjacent to exactly two other crumbs. The ant is
placed at one end of the path and on a square containing a crumb. For example, the
figure below shows a situation in which the ant faces North, and there is a trail of
food leading approximately Southeast. The ant has already eaten the crumb upon
which it was initially placed.
The ant can only smell food directly in front of it. The ant can only remember
a small number of things, and what it remembers after any move only depends on
what it remembered and smelled immediately before the move. Based on smell and
memory, the ant may choose to move forward one square, or it may turn right or
left. It eats a crumb when it lands on it.
The above scenario can be nicely modelled as a state machine in which each state
is a pair consisting of the “ant’s memory” and “everything else”—for example,
information about where things are on the grid. Work out the details of such a
model state machine; design the ant-memory part of the state machine so the ant
will eat all the crumbs on any finite path at which it starts and then signal when it
is done. Be sure to clearly describe the possible states, transitions, and inputs and
outputs (if any) in your model. Briefly explain why your ant will eat all the crumbs.
Note that the last transition is a self-loop; the ant signals done for eternity. One
could also add another end state so that the ant signals done only once.
“mcs” — 2017/6/5 — 19:42 — page 191 — #199
6.4. The Stable Marriage Problem
191
Problem 6.5.
Suppose that you have a regular deck of cards arranged as follows, from top to
bottom:
A~ 2~ : : : K~ A 2 : : : K A| 2| : : : K| A} 2} : : : K}
Only two operations on the deck are allowed: inshuffling and outshuffling. In
both, you begin by cutting the deck exactly in half, taking the top half into your
right-hand and the bottom into your left. Then you shuffle the two halves together
so that the cards are perfectly interlaced; that is, the shuffled deck consists of one
card from the left, one from the right, one from the left, one from the right, etc. The
top card in the shuffled deck comes from the right-hand in an outshuffle and from
the left-hand in an inshuffle.
(a) Model this problem as a state machine.
(b) Use the Invariant Principle to prove that you cannot make the entire first half
of the deck black through a sequence of inshuffles and outshuffles.
Note: Discovering a suitable invariant can be difficult! This is the part of a
correctness proof that generally requires some insight, and there is no simple recipe
for finding invariants. A standard initial approach is to identify a bunch of reachable
states and then look for a pattern—some feature that they all share.
Problem 6.6.
Prove that the fast exponentiation state machine of Section 6.3.1 will halt after
dlog2 ne C 1
(6.3)
transitions starting from any state where the value of z is n 2 ZC .
Hint: Strong induction.
Class Problems
Problem 6.7.
In this problem you will establish a basic property of a puzzle toy called the Fifteen
Puzzle using the method of invariants. The Fifteen Puzzle consists of sliding square
tiles numbered 1; : : : ; 15 held in a 4 4 frame with one empty square. Any tile
adjacent to the empty square can slide into it.
“mcs” — 2017/6/5 — 19:42 — page 192 — #200
192
Chapter 6
State Machines
The standard initial position is
1 2 3 4
5 6 7 8
9 10 11 12
13 14 15
We would like to reach the target position (known in the oldest author’s youth as
“the impossible”):
15 14 13 12
11 10 9 8
7 6 5 4
3 2 1
A state machine model of the puzzle has states consisting of a 4 4 matrix with
16 entries consisting of the integers 1; : : : ; 15 as well as one “empty” entry—like
each of the two arrays above.
The state transitions correspond to exchanging the empty square and an adjacent
numbered tile. For example, an empty at position .2; 2/ can exchange position with
tile above it, namely, at position .1; 2/:
n1 n2 n3 n4
n5
n6 n7
n8 n9 n10 n11
n12 n13 n14 n15
n1
n3 n4
n5 n2 n6 n7
!
n8 n9 n10 n11
n12 n13 n14 n15
We will use the invariant method to prove that there is no way to reach the target
state starting from the initial state.
We begin by noting that a state can also be represented as a pair consisting of
two things:
1. a list of the numbers 1; : : : ; 15 in the order in which they appear—reading
rows left-to-right from the top row down, ignoring the empty square, and
2. the coordinates of the empty square—where the upper left square has coordinates .1; 1/, the lower right .4; 4/.
(a) Write out the “list” representation of the start state and the “impossible” state.
Let L be a list of the numbers 1; : : : ; 15 in some order. A pair of integers is
an out-of-order pair in L when the first element of the pair both comes earlier in
the list and is larger, than the second element of the pair. For example, the list
1; 2; 4; 5; 3 has two out-of-order pairs: (4,3) and (5,3). The increasing list 1; 2 : : : n
has no out-of-order pairs.
“mcs” — 2017/6/5 — 19:42 — page 193 — #201
6.4. The Stable Marriage Problem
193
Let a state S be a pair .L; .i; j // described above. We define the parity of S to
be 0 or 1 depending on whether the sum of the number of out-of-order pairs in L
and the row-number of the empty square is even or odd. that is
(
0 if p.L/ C i is even;
parity.S / WWD
1 otherwise:
(b) Verify that the parity of the start state and the target state are different.
(c) Show that the parity of a state is preserved under transitions. Conclude that
“the impossible” is impossible to reach.
By the way, if two states have the same parity, then in fact there is a way to get
from one to the other. If you like puzzles, you’ll enjoy working this out on your
own.
Problem 6.8.
The Massachusetts Turnpike Authority is concerned about the integrity of the new
Zakim bridge. Their consulting architect has warned that the bridge may collapse
if more than 1000 cars are on it at the same time. The Authority has also been
warned by their traffic consultants that the rate of accidents from cars speeding
across bridges has been increasing.
Both to lighten traffic and to discourage speeding, the Authority has decided to
make the bridge one-way and to put tolls at both ends of the bridge (don’t laugh, this
is Massachusetts). So cars will pay tolls both on entering and exiting the bridge, but
the tolls will be different. In particular, a car will pay $3 to enter onto the bridge and
will pay $2 to exit. To be sure that there are never too many cars on the bridge, the
Authority will let a car onto the bridge only if the difference between the amount
of money currently at the entry toll booth and the amount at the exit toll booth is
strictly less than a certain threshold amount of $T0 .
The consultants have decided to model this scenario with a state machine whose
states are triples .A; B; C / of nonnegative integers, where
A is an amount of money at the entry booth,
B is an amount of money at the exit booth, and
C is a number of cars on the bridge.
Any state with C > 1000 is called a collapsed state, which the Authority dearly
hopes to avoid. There will be no transition out of a collapsed state.
“mcs” — 2017/6/5 — 19:42 — page 194 — #202
194
Chapter 6
State Machines
Since the toll booth collectors may need to start off with some amount of money
in order to make change, and there may also be some number of “official” cars
already on the bridge when it is opened to the public, the consultants must be ready
to analyze the system started at any uncollapsed state. So let A0 be the initial
number of dollars at the entrance toll booth, B0 the initial number of dollars at the
exit toll booth, and C0 1000 the number of official cars on the bridge when it is
opened. You should assume that even official cars pay tolls on exiting or entering
the bridge after the bridge is opened.
(a) Give a mathematical model of the Authority’s system for letting cars on and off
the bridge by specifying a transition relation between states of the form .A; B; C /
above.
(b) Characterize each of the following derived variables
A; B; A C B; A
B; 3C
A; 2A
3B; B C 3C; 2A
3B
6C; 2A
2B
3C
as one of the following
constant
strictly increasing
strictly decreasing
weakly increasing but not constant
weakly decreasing but not constant
none of the above
C
SI
SD
WI
WD
N
and briefly explain your reasoning.
The Authority has asked their engineering consultants to determine T and to
verify that this policy will keep the number of cars from exceeding 1000.
The consultants reason that if C0 is the number of official cars on the bridge
when it is opened, then an additional 1000 C0 cars can be allowed on the bridge.
So as long as A B has not increased by 3.1000 C0 /, there shouldn’t more than
1000 cars on the bridge. So they recommend defining
T0 WWD 3.1000
C0 / C .A0
B0 /;
(6.4)
where A0 is the initial number of dollars at the entrance toll booth, B0 is the initial
number of dollars at the exit toll booth.
(c) Use the results of part (b) to define a simple predicate P on states of the transition system which is satisfied by the start state—that is P .A0 ; B0 ; C0 / holds—is
not satisfied by any collapsed state, and is a preserved invariant of the system. Explain why your P has these properties. Conclude that the traffic won’t cause the
bridge to collapse.
“mcs” — 2017/6/5 — 19:42 — page 195 — #203
6.4. The Stable Marriage Problem
195
(d) A clever MIT intern working for the Turnpike Authority agrees that the Turnpike’s bridge management policy will be safe: the bridge will not collapse. But she
warns her boss that the policy will lead to deadlock—a situation where traffic can’t
move on the bridge even though the bridge has not collapsed.
Explain more precisely in terms of system transitions what the intern means, and
briefly, but clearly, justify her claim.
Problem 6.9.
Start with 102 coins on a table, 98 showing heads and 4 showing tails. There are
two ways to change the coins:
(i) flip over any ten coins, or
(ii) let n be the number of heads showing. Place n C 1 additional coins, all
showing tails, on the table.
For example, you might begin by flipping nine heads and one tail, yielding 90
heads and 12 tails, then add 91 tails, yielding 90 heads and 103 tails.
(a) Model this situation as a state machine, carefully defining the set of states, the
start state, and the possible state transitions.
(b) Explain how to reach a state with exactly one tail showing.
(c) Define the following derived variables:
C WWD the number of coins on the table; H WWD the number of heads;
T WWD the number of tails;
C2 WWD remainder.C =2/;
H2 WWD remainder.H=2/;
T2 WWD remainder.T =2/:
Which of these variables is
1.
2.
3.
4.
5.
strictly increasing
weakly increasing
strictly decreasing
weakly decreasing
constant
(d) Prove that it is not possible to reach a state in which there is exactly one head
showing.
“mcs” — 2017/6/5 — 19:42 — page 196 — #204
196
Chapter 6
State Machines
Problem 6.10.
A classroom is designed so students sit in a square arrangement. An outbreak of
beaver flu sometimes infects students in the class; beaver flu is a rare variant of bird
flu that lasts forever, with symptoms including a yearning for more quizzes and the
thrill of late night problem set sessions.
Here is an illustration of a 66-seat classroom with seats represented by squares.
The locations of infected students are marked with an asterisk.
Outbreaks of infection spread rapidly step by step. A student is infected after a
step if either
the student was infected at the previous step (since beaver flu lasts forever),
or
the student was adjacent to at least two already-infected students at the previous step.
Here adjacent means the students’ individual squares share an edge (front, back,
left or right); they are not adjacent if they only share a corner point. So each student
is adjacent to 2, 3 or 4 others.
In the example, the infection spreads as shown below.
)
)
In this example, over the next few time-steps, all the students in class become
infected.
Theorem. If fewer than n students among those in an nn arrangment are initially
infected in a flu outbreak, then there will be at least one student who never gets
infected in this outbreak, even if students attend all the lectures.
“mcs” — 2017/6/5 — 19:42 — page 197 — #205
6.4. The Stable Marriage Problem
197
Prove this theorem.
Hint: Think of the state of an outbreak as an n n square above, with asterisks
indicating infection. The rules for the spread of infection then define the transitions
of a state machine. Find a weakly decreasing derived variable that leads to a proof
of this theorem.
Exam Problems
Problem 6.11.
Token replacing-1-2 is a single player game using a set of tokens, each colored
black or white. Except for color, the tokens are indistinguishable. In each move,
a player can replace one black token with two white tokens, or replace one white
token with two black tokens.
We can model this game as a state machine whose states are pairs .nb ; nw / where
nb 0 equals the number of black tokens, and nw 0 equals the number of white
tokens.
(a) List the numbers of the following predicates that are preserved invariants.
nb C nw rem.nb C nw ; 3/ ¤ 2
(6.5)
nw
nb rem.nw
nb ; 3/ D 2
(6.6)
nb
nw rem.nb
n2 ; 3/ D 2
(6.7)
nb C nw > 5
(6.8)
nb C nw < 5
(6.9)
Now assume the game starts with a single black token, that is, the start state is
.1; 0/.
(b) List the numbers of the predicates above are true for all reachable states:
(c) Define the predicate T .nb ; nw / by the rule:
T .nb ; nw / WWD rem.nw
nb ; 3/ D 2:
We will now prove the following:
Claim. If T .nb ; nw /, then state .nb ; nw / is reachable.
“mcs” — 2017/6/5 — 19:42 — page 198 — #206
198
Chapter 6
State Machines
Note that this claim is different from the claim that T is a preserved invariant.
The proof of the Claim will be by induction in n using induction hypothesis P .n/WWD
8.nb ; nw /: Œ.nb C nw D n/ AND T .nb ; nw / IMPLIES .nb ; nw / is reachable:
The base cases will be when n 2.
Assuming that the base cases have been verified, complete the Inductive
Step.
Now verify the Base Cases: P .n/ for n 2.
Problem 6.12.
Token Switching is a process for updating a set of black and white tokens. The
process starts with a single black token. At each step,
(i) one black token can be replaced with two white tokens, or
(ii) if the numbers of white and black tokens are not the same, the colors of all
the tokens can be switched: all the black tokens become white, and the white
tokens become black.
We can model Token Switching as a state machine whose states are pairs .b; w/
of nonnegative integers, where b equals the number of black tokens, and w equals
the number of white tokens. So the start state is .1; 0/.
(a) Indicate which of the following states can be reached from the start state in
exactly two steps:
.0; 0/; .1; 0/; .0; 1/; .1; 1/; .0; 2/; .2; 0/; .2; 1/; .1; 2/; .0; 3/; .3; 0/
(b) Define the predicate F .b; w/ by the rule:
F .b; w/ WWD .b
w/is not a multiple of 3:
Prove the following
Claim. If F .b; w/, then state .b; w/ is reachable from the start state.
(c) Explain why state .116
7777
88
; 510 / is not a reachable state.
Hint: Do not assume F is a preserved invariant without proving it.
“mcs” — 2017/6/5 — 19:42 — page 199 — #207
6.4. The Stable Marriage Problem
199
Problem 6.13.
Token replacing-1-3 is a single player game using a set of tokens, each colored
black or white. In each move, a player can replace a black token with three white
tokens, or replace a white token with three black tokens. We can model this game
as a state machine whose states are pairs .b; w/ of nonnegative integers, where b is
the number of black tokens and w the number of white ones.
The game has two possible start states: .5; 4/ or .4; 3/.
We call a state .b; w/ eligible when
rem.b
w; 4/ D 1; AND
(6.10)
minfb; wg 3:
(6.11)
This problem examines the connection between eligible states and states that are
reachable from either of the possible start states.
(a) Give an example of a reachable state that is not eligible.
(b) Show that the derived variable b C w is strictly increasing. Conclude that state
.3; 2/ is not reachable.
(c) Suppose .b; w/ is eligible and b 6. Verify that .b
3; w C 1/ is eligible.
For the rest of the problem, you may—and should—assume the following Fact:
Fact. If maxfb; wg 5 and .b; w/ is eligible, then .b; w/ is reachable.
(This is easy to verify since there are only nine states with b; w 2 f3; 4; 5g, but
don’t waste time doing this.)
(d) Define the predicate P .n/ to be:
8.b; w/:Œb C w D n AND .b; w/ is eligible IMPLIES .b; w/ is reachable:
Prove that P .n
1/ IMPLIES P .n C 1/ for all n 1.
(e) Conclude that all eligible states are reachable.
(f) Prove that .47 C 1; 45 C 2/ is not reachable.
(g) Verify that rem.3b w; 8/ is a derived variable that is constant. Conclude that
no state is reachable from both start states.
“mcs” — 2017/6/5 — 19:42 — page 200 — #208
200
Chapter 6
State Machines
Problem 6.14.
There is a bucket containing more blue balls than red balls. As long as there are
more blues than reds, any one of the following rules may be applied to add and/or
remove balls from the bucket:
(i) Add a red ball.
(ii) Remove a blue ball.
(iii) Add two reds and one blue.
(iv) Remove two blues and one red.
(a) Starting with 10 reds and 16 blues, what is the largest number of balls the
bucket will contain by applying these rules?
Let b be the number of blue balls and r be the number of red balls in the bucket
at any given time.
(b) Prove that b r 0 is a preserved invariant of the process of adding and
removing balls according to rules (i)–(iv).
(c) Prove that no matter how many balls the bucket contains, repeatedly applying
rules (i)–(iv) will eventually lead to a state where no further rule can be applied.
Problem 6.15.
The following problem is a twist on the Fifteen-Puzzle considered earlier in Problem 6.7.
Let A be a sequence consisting of the numbers 1; : : : ; n in some order. A pair
of integers in A is called an out-of-order pair when the first element of the pair
both comes earlier in the sequence, and is larger, than the second element of the
pair. For example, the sequence .1; 2; 4; 5; 3/ has two out-of-order pairs: .4; 3/
and .5; 3/. We let t .A/ equal the number of out-of-order pairs in A. For example,
t ..1; 2; 4; 5; 3// D 2.
The elements in A can be rearranged using the Rotate-Triple operation, in which
three consecutive elements of A are rotated to move the smallest of them to be first.
For example, in the sequence .2; 4; 1; 5; 3/, the Rotate-Triple operation could
rotate the consecutive numbers 4; 1; 5, into 1; 5; 4 so that
.2; 4; 1; 5; 3/ ! .2; 1; 5; 4; 3/:
The Rotate-Triple could also rotate the consecutive numbers 2; 4; 1 into 1; 2; 4
so that
.2; 4; 1; 5; 3/ ! .1; 2; 4; 5; 3/:
“mcs” — 2017/6/5 — 19:42 — page 201 — #209
6.4. The Stable Marriage Problem
201
We can think of a sequence A as a state of a state machine whose transitions
correspond to possible applications of the Rotate-Triple operation.
(a) Argue that the derived variable t is weakly decreasing.
(b) Prove that having an even number of out-of-order pairs is a preserved invariant
of this machine.
(c) Starting with
S WWD .2014; 2013; 2012; : : : ; 2; 1/;
explain why it is impossible to reach
T WWD .1; 2; : : : ; 2012; 2013; 2014/:
Problems for Section 6.4
Practice Problems
Problem 6.16.
Four Students want separate assignments to four VI-A Companies. Here are their
preference rankings:
Student
Albert:
Sarah:
Tasha:
Elizabeth:
Company
AT&T:
Bellcore:
HP:
Draper:
Companies
HP, Bellcore, AT&T, Draper
AT&T, Bellcore, Draper, HP
HP, Draper, AT&T, Bellcore
Draper, AT&T, Bellcore, HP
Students
Elizabeth, Albert, Tasha, Sarah
Tasha, Sarah, Albert, Elizabeth
Elizabeth, Tasha, Albert, Sarah
Sarah, Elizabeth, Tasha, Albert
(a) Use the Mating Ritual to find two stable assignments of Students to Companies.
(b) Describe a simple procedure to determine whether any given stable marriage
problem has a unique solution, that is, only one possible stable matching. Briefly
explain why it works.
“mcs” — 2017/6/5 — 19:42 — page 202 — #210
202
Chapter 6
State Machines
Problem 6.17.
Suppose that Harry is one of the boys and Alice is one of the girls in the Mating
Ritual. Which of the properties below are preserved invariants? Why?
a. Alice is the only girl on Harry’s list.
b. There is a girl who does not have any boys serenading her.
c. If Alice is not on Harry’s list, then Alice has a suitor that she prefers to Harry.
d. Alice is crossed off Harry’s list, and Harry prefers Alice to anyone he is
serenading.
e. If Alice is on Harry’s list, then she prefers Harry to any suitor she has.
Problem 6.18.
Prove that whatever the marriage preferences among the men and women, every
man is the pessimal husband of his optimal wife.
Hint: Follows directly from the definition of “rogue couple.”
Problem 6.19.
In the Mating Ritual for stable marriages between an equal number of boys and
girls, explain why there must be a girl to whom no boy proposes (serenades) until
the last day.
Class Problems
Problem 6.20.
The preferences among 4 boys and 4 girls are partially specified in the following
table:
B1:
B2:
B3:
B4:
G1:
G2:
G3:
G4:
G1
G2
–
–
B2
B1
–
–
G2
G1
–
–
B1
B2
–
–
–
–
G4
G3
–
–
B3
B4
–
–
G3
G4
–
–
B4
B3
“mcs” — 2017/6/5 — 19:42 — page 203 — #211
6.4. The Stable Marriage Problem
203
(a) Verify that
.B1; G1/; .B2; G2/; .B3; G3/; .B4; G4/
will be a stable matching whatever the unspecified preferences may be.
(b) Explain why the stable matching above is neither boy-optimal nor boy-pessimal
and so will not be an outcome of the Mating Ritual.
(c) Describe how to define a set of marriage preferences among n boys and n girls
which have at least 2n=2 stable assignments.
Hint: Arrange the boys into a list of n=2 pairs, and likewise arrange the girls into
a list of n=2 pairs of girls. Choose preferences so that the kth pair of boys ranks
the kth pair of girls just below the previous pairs of girls, and likewise for the kth
pair of girls. Within the kth pairs, make sure each boy’s first choice girl in the pair
prefers the other boy in the pair.
Problem 6.21.
The Mating Ritual of Section 6.4.1 for finding stable marriages works even when
the numbers of men and women are not equal. As before, a set of (monogamous)
marriages between men and women is called stable when it has no “rogue couples.”
(a) Extend the definition of rogue couple so it covers the case of unmarried men
and women. Verify that in a stable set of marriages, either all the men are married
or all the women are married.
(b) Explain why even in the case of unequal numbers of men and women, applying
the Mating Ritual will yield a stable matching.
Homework Problems
Problem 6.22.
Suppose we want to assign pairs of “buddies,” who may be of the sex, where each
person has a preference rank for who they would like to be buddies with. For the
preference ranking given in Figure 6.5, show that there is no stable buddy assignment. In this figure Mergatroid’s preferences aren’t shown because they don’t even
matter.
Problem 6.23.
The most famous application of stable matching was in assigning graduating medical students to hospital residencies. Each hospital has a preference ranking of
“mcs” — 2017/6/5 — 19:42 — page 204 — #212
204
Chapter 6
State Machines
Alex
2 1
3
Robin
2
1
1 3
3 2
Bobby Joe
Mergatroid
Figure 6.5 Some preferences with no stable buddy matching.
students, and each student has a preference ranking of hospitals, but unlike finding
stable marriages between an equal number of boys and girls, hospitals generally
have differing numbers of available residencies, and the total number of residencies may not equal the number of graduating students.
Explain how to adapt the Stable Matching problem with an equal number of boys
and girls to this more general situation. In particular, modify the definition of stable
matching so it applies in this situation, and explain how to adapt the Mating Ritual
to handle it.
Problem 6.24.
Give an example of a stable matching between 3 boys and 3 girls where no person
gets their first choice. Briefly explain why your matching is stable. Can your
matching be obtained from the Mating Ritual or the Ritual with boys and girls
reversed?
Problem 6.25.
In a stable matching between an equal number of boys and girls produced by the
Mating Ritual, call a person lucky if they are matched up with someone in the top
half of their preference list. Prove that there must be at least one lucky person.
Hint: The average number of times a boy gets rejected by girls.
Problem 6.26.
Suppose there are two stable sets of marriages. So each man has a first wife and a
“mcs” — 2017/6/5 — 19:42 — page 205 — #213
6.4. The Stable Marriage Problem
205
second wife , and likewise each woman has a first husband and a second husband.
Someone in a given marriage is a winner when they prefer their current spouse
to their other spouse, and they are a loser when they prefer their other spouse to
their current spouse. (If someone has the same spouse in both of their marriages,
then they will be neither a winner nor a loser.)
We will show that
In each of the marriages, someone is a winner iff their spouse is a loser.
(WL)
(a) The left to right direction of (WL) is equivalent to the assertion that married
partners cannot both be winners. Explain why this follows directly from the definition of rogue couple.
The right to left direction of (WL) is equivalent to the assertion that a married
couple cannot both be losers. This will follow by comparing the number of winners
and losers among the marriages.
(b) Explain why the number of winners must equal the number of losers among
the two sets of marriages.
(c) Complete the proof of (WL) by showing that if some married couple were both
losers, then there must be another couple who were both winners.
(d) Conclude that in a stable set of marriages, someone’s spouse is optimal iff they
are pessimal for their spouse.
Problem 6.27.
Suppose there are two stable sets of marriages, a first set and a second set. So
each man has a first wife and a second wife (they may be the same), and likewise
each woman has a first husband and a second husband. We can form a third set
of marriages by matching each man with the wife he prefers among his first and
second wives.
(a) Prove that this third set of marriages is an exact matching: no woman is married to two men.
(b) Prove that this third marriage set is stable.
Hint: You may assume the following fact from Problem 6.26.
In every marriage, someone is a winner iff their spouse is a loser.
(SL)
“mcs” — 2017/6/5 — 19:42 — page 206 — #214
206
Chapter 6
State Machines
Problem 6.28.
A state machine has commuting transitions if for any states p; q; r
.p ! q AND p ! r/ IMPLIES 9t: q ! t AND r ! t:
The state machine is confluent if
.p ! q AND p ! r/ IMPLIES 9t: q ! t AND r ! t:
(a) Prove that if a state machine has commuting transitions, then it is confluent.
Hint: By induction on the number of moves from p to q plus the number from p
to r.
(b) A final state of a state machine is one from which no transition is possible.
Explain why, if a state machine is confluent, then at most one final state is reachable
from the start state.
Problem 6.29.
According to the day-by-day description of the Mating Ritual of Section 6.4.1, at
the end of each day, every man’s list is updated to remove the name of the woman
he who rejected him. But it’s easier, and more flexible, simply to let one women
reject one suitor at a time.
In particular, the states of this Flexible Mating Ritual state machine will be the
same as for the day-by-day Ritual: a state will be a list, for each man, of the women
who have not rejected him. But now a transition will be to choose two men who
are serenading the same woman—that is, who have the same woman at the top
of their current lists—and then have the woman reject whichever of the two she
likes less. So the only change in state is that the name of the serenaded woman
gets deleted from the top of the list of the man she liked less among two of her
serenaders—everything else stays the same.
It’s a worthwhile review to verify that the same preserved invariants used to establish the properties of the Mating Ritual will apply to the Flexible Mating Ritual.
This ensures that the Flexible Ritual will also terminate with a stable set of marriages.
But now a new issue arises: we know that there can be many sets of possible
sets of stable marriages for the same set of men/women preferences. So it seems
possible that the Flexible Ritual might terminate with different stable marriage sets,
depending on which choice of transition was made at each state. But this does
not happen: the Flexible Ritual will always terminate with the same set of stable
marriages as the day-by-day Ritual.
“mcs” — 2017/6/5 — 19:42 — page 207 — #215
6.4. The Stable Marriage Problem
207
To prove this, we begin with a definition: a state machine has commuting transitions if for any states p; q; r,
.p ! q AND p ! r/ IMPLIES 9t: q ! t AND r ! t:
(a) Verify that the Flexible Mating Ritual has commuting transitions.
(b) Now conclude from Problem 6.28 that the Flexible Mating Ritual always terminate with the same set of stable marriages as the day-by-day Ritual.
Exam Problems
Problem 6.30.
Four unfortunate children want to be adopted by four foster families of ill repute.
A child can only be adopted by one family, and a family can only adopt one child.
Here are their preference rankings (most-favored to least-favored):
Child
Bottlecap:
Lucy:
Dingdong:
Zippy:
Families
Hatfields, McCoys, Grinches, Scrooges
Grinches, Scrooges, McCoys, Hatfields
Hatfields, Scrooges, Grinches, McCoys
McCoys, Grinches, Scrooges, Hatfields
Family
Grinches:
Hatfields:
Scrooges:
McCoys:
Children
Zippy, Dingdong, Bottlecap, Lucy
Zippy, Bottlecap, Dingdong, Lucy
Bottlecap, Lucy, Dingdong, Zippy
Lucy, Zippy, Bottlecap, Dingdong
(a) Exhibit two different stable matching of Children and Families.
Family
Grinches:
Hatfields:
Scrooges:
McCoys:
Child in 1st match
Child in 2nd match
(b) Examine the matchings from part a, and explain why these matchings are the
only two possible stable matchings between Children and Families.
Hint: In general, there may be many more than two stable matchings for the same
set of preferences.
“mcs” — 2017/6/5 — 19:42 — page 208 — #216
208
Chapter 6
State Machines
Problem 6.31.
The Mating Ritual 6.4.1 for finding stable marriages works without change when
there are at least as many, and possibly more, men than women. You may assume
this. So the Ritual ends with all the women married and no rogue couples for these
marriages, where an unmarried man and a married woman who prefers him to her
spouse is also considered to be a “rogue couple.”
Let Alice be one of the women, and Bob be one of the men. Indicate which of
the properties below that are preserved invariants of the Mating Ritual 6.4 when
there are at least as many men as women. Briefly explain your answers.
(a) Alice has a suitor (man who is serenading her) whom she prefers to Bob.
(b) Alice is the only woman on Bob’s list.
(c) Alice has no suitor.
(d) Bob prefers Alice to the women he is serenading.
(e) Bob is serenading Alice.
(f) Bob is not serenading Alice.
(g) Bob’s list of women to serenade is empty.
Problem 6.32.
We want a stable matching between n boys and n girls for a positive integer n.
(a) Explain how to define preference rankings for the boys and the girls that allow
only one possible stable matching. Briefly justify your answer.
(b) Mark each of the following predicates about the Stable Marriage Ritual P if it
is a Preserved Invariant, N if it is not, and “U” if you are very unsure. “Bob’s list”
refers to the list of the women he has not crossed off.
(i) Alice is not on Bob’s list.
(ii) No girl is on Bob’s list.
(iii) Bob is the only boy serenading Alice.
(iv) Bob has fewer than 5 girls on his list.
(v) Bob prefers Alice to his favorite remaining girl.
(vi) Alice prefers her favorite current suitor to Bob.
(vii) Bob is serenading his optimal spouse.
“mcs” — 2017/6/5 — 19:42 — page 209 — #217
6.4. The Stable Marriage Problem
(viii) Bob is serenading his pessimal spouse.
(ix) Alice’s optimal spouse is serenading her.
(x) Alice’s pessimal spouse is serenading her.
209
“mcs” — 2017/6/5 — 19:42 — page 210 — #218
“mcs” — 2017/6/5 — 19:42 — page 211 — #219
7
Recursive Data Types
Recursive data types play a central role in programming, and induction is really all
about them.
Recursive data types are specified by recursive definitions, which say how to
construct new data elements from previous ones. Along with each recursive data
type there are recursive definitions of properties or functions on the data type. Most
importantly, based on a recursive definition, there is a structural induction method
for proving that all data of the given type have some property.
This chapter examines a few examples of recursive data types and recursively
defined functions on them:
strings of characters,
“balanced” strings of brackets,
the nonnegative integers, and
arithmetic expressions.
two-player games with perfect information.
7.1
Recursive Definitions and Structural Induction
We’ll start off illustrating recursive definitions and proofs using the example of
character strings. Normally we’d take strings of characters for granted, but it’s
informative to treat them as a recursive data type. In particular, strings are a nice
first example because you will see recursive definitions of things that are easy to
understand, or that you already know, so you can focus on how the definitions work
without having to figure out what they are supposed to mean.
Definitions of recursive data types have two parts:
Base case(s) specifying that some known mathematical elements are in the
data type, and
Constructor case(s) that specify how to construct new data elements from
previously constructed elements or from base elements.
The definition of strings over a given character set A follows this pattern:
“mcs” — 2017/6/5 — 19:42 — page 212 — #220
212
Chapter 7
Recursive Data Types
Definition 7.1.1. Let A be a nonempty set called an alphabet, whose elements are
referred to as characters (also called letters, symbols, or digits). The recursive data
type A of strings over alphabet A is defined as follows:
Base case: the empty string is in A .
Constructor case: If a 2 A and s 2 A , then the pair ha; si 2 A .
So f0; 1g are the binary strings.
The usual way to treat binary strings is as sequences of 0’s and 1’s. For example,
we have identified the length-4 binary string 1011 as a sequence of bits, the 4-tuple
.1; 0; 1; 1/. But according to the recursive Definition 7.1.1, this string would be
represented by nested pairs, namely
h1; h0; h1; h1; iiii :
These nested pairs are definitely cumbersome and may also seem bizarre, but they
actually reflect the way that such lists of characters would be represented in programming languages like Scheme or Python, where ha; si would correspond to
cons.a; s/.
Notice that we haven’t said exactly how the empty string is represented. It really
doesn’t matter, as long as we can recognize the empty string and not confuse it with
any nonempty string.
Continuing the recursive approach, let’s define the length of a string.
Definition 7.1.2. The length jsj of a string s is defined recursively based on Definition 7.1.1.
Base case: jj WWD 0.
Constructor case: j ha; si j WWD 1 C jsj.
This definition of length follows a standard pattern: functions on recursive data
types can be defined recursively using the same cases as the data type definition.
Specifically, to define a function f on a recursive data type, define the value of f
for the base cases of the data type definition, then define the value of f in each
constructor case in terms of the values of f on the component data items.
Let’s do another example: the concatenation s t of the strings s and t is the
string consisting of the letters of s followed by the letters of t. This is a perfectly clear mathematical definition of concatenation (except maybe for what to do
with the empty string), and in terms of Scheme/Python lists, s t would be the list
append.s; t /. Here’s a recursive definition of concatenation.
“mcs” — 2017/6/5 — 19:42 — page 213 — #221
7.1. Recursive Definitions and Structural Induction
213
Definition 7.1.3. The concatenation s t of the strings s; t 2 A is defined recursively based on Definition 7.1.1:
Base case:
t WWD t:
Constructor case:
ha; si t WWD ha; s ti :
7.1.1
Structural Induction
Structural induction is a method for proving that all the elements of a recursively
defined data type have some property. A structural induction proof has two parts
corresponding to the recursive definition:
Prove that each base case element has the property.
Prove that each constructor case element has the property, when the constructor is applied to elements that have the property.
For example, in the base case of the definition of concatenation 7.1.3, we defined
concatenation so the empty string was a “left identity,” namely, s WWD s. We want
the empty string also to be “right identity,” namely, s D s. Being a right identity
is not part of Definition 7.1.3, but we can prove it easily by structural induction:
Lemma 7.1.4.
sDs
for all s 2 A .
Proof. The proof is by structural induction on the recursive definition 7.1.3 of concatenation. The induction hypothesis will be
P .s/ WWD Œs D s:
Base case: (s D ).
sD
D
D s:
( is a left identity by Def 7.1.3)
“mcs” — 2017/6/5 — 19:42 — page 214 — #222
214
Chapter 7
Recursive Data Types
Constructor case: (s D a t ).
s D .a t / D a .t /
(Constructor case of Def 7.1.3)
Dat
by induction hypothesis P .t /
D s:
So P .s/ holds. This completes the proof of the constructor case, and we conclude
by structural induction that equation (7.1.4) holds for all s 2 A .
We can also verify properties of recursive functions by structural induction on
their definitions. For example, let’s verify the familiar fact that the length of the
concatenation of two strings is the sum of their lengths:
Lemma.
js t j D jsj C jtj
for all s; t 2 A .
Proof. By structural induction on the definition of s 2 A . The induction hypothesis is
P .s/ WWD 8t 2 A : js t j D jsj C jtj:
Base case (s D ):
js tj D j tj
D jt j
(base case of Def 7.1.3 of concatenation)
D 0 C jt j
D jsj C jt j
(Def of jj):
Constructor case: (s WWD ha; ri).
js tj D j ha; ri tj
D j ha; r t i j
D 1 C jr t j
D 1 C .jrj C jtj/
(constructor case of Def of concat)
(constructor case of def length)
(ind. hyp. P .r/)
D .1 C jrj/ C jtj
D j ha; ri j C jtj
(constructor case, def of length)
D jsj C jtj:
This proves that P .s/ holds, completing the constructor case. By structural induction, we conclude that P .s/ holds for all strings s 2 A .
“mcs” — 2017/6/5 — 19:42 — page 215 — #223
7.2. Strings of Matched Brackets
215
These proofs illustrate the general principle:
The Principle of Structural Induction.
Let P be a predicate on a recursively defined data type R. If
P .b/ is true for each base case element b 2 R, and
for all two-argument constructors c,
ŒP .r/ AND P .s/ IMPLIES P .c.r; s//
for all r; s 2 R,
and likewise for all constructors taking other numbers of arguments,
then
P .r/ is true for all r 2 R:
7.2
Strings of Matched Brackets
Let f] ; [ g be the set of all strings of square brackets. For example, the following
two strings are in f] ; [ g :
[]][[[[[]]
and [ [ [ ] ] [ ] ] [ ]
(7.1)
A string s 2 f] ; [ g is called a matched string if its brackets “match up” in
the usual way. For example, the left-hand string above is not matched because its
second right bracket does not have a matching left bracket. The string on the right
is matched.
We’re going to examine several different ways to define and prove properties
of matched strings using recursively defined sets and functions. These properties
are pretty straightforward, and you might wonder whether they have any particular
relevance in computer science. The honest answer is “not much relevance any
more.” The reason for this is one of the great successes of computer science, as
explained in the text box below.
“mcs” — 2017/6/5 — 19:42 — page 216 — #224
216
Chapter 7
Recursive Data Types
Expression Parsing
During the early development of computer science in the 1950’s and 60’s, creation
of effective programming language compilers was a central concern. A key aspect
in processing a program for compilation was expression parsing. One significant
problem was to take an expression like
x C y z2 y C 7
and put in the brackets that determined how it should be evaluated—should it be
ŒŒx C y z 2 y C 7; or;
x C Œy z 2 Œy C 7; or;
Œx C Œy z 2  Œy C 7; or : : :‹
The Turing award (the “Nobel Prize” of computer science) was ultimately bestowed on Robert W. Floyd, for, among other things, discovering simple procedures that would insert the brackets properly.
In the 70’s and 80’s, this parsing technology was packaged into high-level
compiler-compilers that automatically generated parsers from expression grammars. This automation of parsing was so effective that the subject no longer
demanded attention. It had largely disappeared from the computer science curriculum by the 1990’s.
The matched strings can be nicely characterized as a recursive data type:
Definition 7.2.1. Recursively define the set RecMatch of strings as follows:
Base case: 2 RecMatch.
Constructor case: If s; t 2 RecMatch, then
[ s ] t 2 RecMatch:
Here [ s ] t refers to the concatenation of strings which would be written in full
as
[ .s .] t //:
From now on, we’ll usually omit the “’s.”
Using this definition, 2 RecMatch by the base case, so letting s D t D in
the constructor case implies
[ ] D [ ] 2 RecMatch:
“mcs” — 2017/6/5 — 19:42 — page 217 — #225
7.2. Strings of Matched Brackets
217
Now,
[ ] [ ] D [ ] [ ] 2 RecMatch
[ [ ] ] D [ [ ] ] 2 RecMatch
[ [ ] ] [ ] 2 RecMatch
(letting s D ; t D [ ] )
(letting s D [ ] ; t D )
(letting s D [ ] ; t D [ ] )
are also strings in RecMatch by repeated applications of the constructor case; and
so on.
It’s pretty obvious that in order for brackets to match, there had better be an equal
number of left and right ones. For further practice, let’s carefully prove this from
the recursive definitions, beginning with a recursive definition of the number #c .s/
of occurrences of the character c 2 A in a string s:
Definition 7.2.2.
Base case: #c ./ WWD 0.
Constructor case:
(
#c .s/
#c .ha; si/ WWD
1 C #c .s/
if a ¤ c;
if a D c:
The following Lemma follows directly by structural induction on Definition 7.2.2.
We’ll leave the proof for practice (Problem 7.9).
Lemma 7.2.3.
#c .s t / D #c .s/ C #c .t /:
Lemma. Every string in RecMatch has an equal number of left and right brackets.
Proof. The proof is by structural induction with induction hypothesis
h
i
P .s/ WWD #[ .s/ D #] .s/ :
Base case: P ./ holds because
#[ ./ D 0 D #] ./
by the base case of Definition 7.2.2 of #c ./.
“mcs” — 2017/6/5 — 19:42 — page 218 — #226
218
Chapter 7
Recursive Data Types
Constructor case: By structural induction hypothesis, we assume P .s/ and P .t /
and must show P .[ s ] t /:
#[ .[ s ] t / D #[ .[ / C #[ .s/ C #[ .] / C #[ .t /
(Lemma 7.2.3)
D 1 C #[ .s/ C 0 C #[ .t /
(def #[ ./)
D 1 C #] .s/ C 0 C #] .t /
(by P .s/ and P .t /)
D 0 C #] .s/ C 1 C #] .t /
D #] .[ / C #] .s/ C #] .] / C #] .t /
D #] .[ s ] t /
(def #] ./)
(Lemma 7.2.3)
This completes the proof of the constructor case. We conclude by structural induction that P .s/ holds for all s 2 RecMatch.
Warning: When a recursive definition of a data type allows the same element
to be constructed in more than one way, the definition is said to be ambiguous.
We were careful to choose an unambiguous definition of RecMatch to ensure that
functions defined recursively on its definition would always be well-defined. Recursively defining a function on an ambiguous data type definition usually will not
work. To illustrate the problem, here’s another definition of the matched strings.
Definition 7.2.4. Define the set, AmbRecMatch f] ; [ g recursively as follows:
Base case: 2 AmbRecMatch,
Constructor cases: if s; t 2 AmbRecMatch, then the strings [ s ] and st are
also in AmbRecMatch.
It’s pretty easy to see that the definition of AmbRecMatch is just another way
to define RecMatch, that is AmbRecMatch D RecMatch (see Problem 7.19). The
definition of AmbRecMatch is arguably easier to understand, but we didn’t use it
because it’s ambiguous, while the trickier definition of RecMatch is unambiguous.
Here’s why this matters. Let’s define the number of operations f .s/ to construct a
matched string s recursively on the definition of s 2 AmbRecMatch:
f ./ WWD 0;
(f base case)
f .[ s ] / WWD 1 C f .s/;
f .st / WWD 1 C f .s/ C f .t /:
(f concat case)
“mcs” — 2017/6/5 — 19:42 — page 219 — #227
7.3. Recursive Functions on Nonnegative Integers
219
This definition may seem ok, but it isn’t: f ./ winds up with two values, and
consequently:
0 D f ./
D f . /
(f base case))
(concat def, base case)
D 1 C f ./ C f ./
(f concat case);
D1C0C0D1
(f base case):
This is definitely not a situation we want to be in!
7.3
Recursive Functions on Nonnegative Integers
The nonnegative integers can be understood as a recursive data type.
Definition 7.3.1. The set N is a data type defined recursively as:
0 2 N.
If n 2 N, then the successor n C 1 of n is in N.
The point here is to make it clear that ordinary induction is simply the special
case of structural induction on the recursive Definition 7.3.1. This also justifies the
familiar recursive definitions of functions on the nonnegative integers.
7.3.1
Some Standard Recursive Functions on N
Example 7.3.2. The factorial function. This function is often written “nŠ.” You will
see a lot of it in later chapters. Here, we’ll use the notation fac.n/:
fac.0/ WWD 1.
fac.n C 1/ WWD .n C 1/ fac.n/ for n 0.
P
Example 7.3.3. Summation notation. Let “S.n/” abbreviate the expression “ niD1 f .i /.”
We can recursively define S.n/ with the rules
S.0/ WWD 0.
S.n C 1/ WWD f .n C 1/ C S.n/ for n 0.
“mcs” — 2017/6/5 — 19:42 — page 220 — #228
220
Chapter 7
7.3.2
Recursive Data Types
Ill-formed Function Definitions
There are some other blunders to watch out for when defining functions recursively.
The main problems come when recursive definitions don’t follow the recursive definition of the underlying data type. Below are some function specifications that resemble good definitions of functions on the nonnegative integers, but really aren’t.
f1 .n/ WWD 2 C f1 .n
1/:
(7.2)
This “definition” has no base case. If some function f1 satisfied (7.2), so would a
function obtained by adding a constant to the value of f1 . So equation (7.2) does
not uniquely define an f1 .
(
f2 .n/ WWD
0;
f2 .n C 1/
if n D 0;
otherwise:
(7.3)
This “definition” has a base case, but still doesn’t uniquely determine f2 . Any
function that is 0 at 0 and constant everywhere else would satisfy the specification,
so (7.3) also does not uniquely define anything.
In a typical programming language, evaluation of f2 .1/ would begin with a recursive call of f2 .2/, which would lead to a recursive call of f2 .3/, . . . with recursive calls continuing without end. This “operational” approach interprets (7.3) as
defining a partial function f2 that is undefined everywhere but 0.
8
ˆ
<0; if n is divisible by 2,
f3 .n/ WWD 1; if n is divisible by 3,
ˆ
:
2; otherwise.
(7.4)
This “definition” is inconsistent: it requires f3 .6/ D 0 and f3 .6/ D 1, so (7.4)
doesn’t define anything.
Mathematicians have been wondering about this function specification, known
as the Collatz conjecture for a while:
8
ˆ
if n 1;
<1;
f4 .n/ WWD f4 .n=2/
(7.5)
if n > 1 is even;
ˆ
:
f4 .3n C 1/ if n > 1 is odd:
For example, f4 .3/ D 1 because
f4 .3/ WWD f4 .10/ WWD f4 .5/ WWD f4 .16/ WWD f4 .8/ WWD f4 .4/ WWD f4 .2/ WWD f4 .1/ WWD 1:
“mcs” — 2017/6/5 — 19:42 — page 221 — #229
7.4. Arithmetic Expressions
221
The constant function equal to 1 will satisfy (7.5), but it’s not known if another
function does as well. The problem is that the third case specifies f4 .n/ in terms
of f4 at arguments larger than n, and so cannot be justified by induction on N. It’s
known that any f4 satisfying (7.5) equals 1 for all n up to over 1018 .
A final example is the Ackermann function, which is an extremely fast-growing
function of two nonnegative arguments. Its inverse is correspondingly slow-growing—
it grows slower than log n, log log n, log log log n, . . . , but it does grow unboundly.
This inverse actually comes up analyzing a useful, highly efficient procedure known
as the Union-Find algorithm. This algorithm was conjectured to run in a number
of steps that grew linearly in the size of its input, but turned out to be “linear”
but with a slow growing coefficient nearly equal to the inverse Ackermann function. This means that pragmatically, Union-Find is linear, since the theoretically
growing coefficient is less than 5 for any input that could conceivably come up.
The Ackermann function can be defined recursively as the function A given by
the following rules:
A.m; n/ D 2n
A.m; n/ D A.m
1; A.m; n
1//
if m D 0 or n 1;
(7.6)
otherwise:
(7.7)
Now these rules are unusual because the definition of A.m; n/ involves an evaluation of A at arguments that may be a lot bigger than m and n. The definitions
of f2 above showed how definitions of function values at small argument values in
terms of larger one can easily lead to nonterminating evaluations. The definition of
the Ackermann function is actually ok, but proving this takes some ingenuity (see
Problem 7.25).
7.4
Arithmetic Expressions
Expression evaluation is a key feature of programming languages, and recognition
of expressions as a recursive data type is a key to understanding how they can be
processed.
To illustrate this approach we’ll work with a toy example: arithmetic expressions
like 3x 2 C 2x C 1 involving only one variable, “x.” We’ll refer to the data type of
such expressions as Aexp. Here is its definition:
Definition 7.4.1.
Base cases:
“mcs” — 2017/6/5 — 19:42 — page 222 — #230
222
Chapter 7
Recursive Data Types
– The variable x is in Aexp.
– The arabic numeral k for any nonnegative integer k is in Aexp.
Constructor cases: If e; f 2 Aexp, then
– [ e + f ] 2 Aexp. The expression [ e + f ] is called a sum. The Aexp’s
e and f are called the components of the sum; they’re also called the
summands.
– [ e * f ] 2 Aexp. The expression [ e * f ] is called a product. The
Aexp’s e and f are called the components of the product; they’re also
called the multiplier and multiplicand.
– - [ e ] 2 Aexp. The expression - [ e ] is called a negative.
Notice that Aexp’s are fully bracketed, and exponents aren’t allowed. So the
Aexp version of the polynomial expression 3x 2 C2x C1 would officially be written
as
[ [ 3 * [ x * x ] ] + [ [ 2 * x ] + 1] ] :
(7.8)
These brackets and ’s clutter up examples, so we’ll often use simpler expressions
like “3x 2 C2xC1” instead of (7.8). But it’s important to recognize that 3x 2 C2xC1
is not an Aexp; it’s an abbreviation for an Aexp.
7.4.1
Evaluation and Substitution with Aexp’s
Evaluating Aexp’s
Since the only variable in an Aexp is x, the value of an Aexp is determined by the
value of x. For example, if the value of x is 3, then the value of 3x 2 C 2x C 1
is 34. In general, given any Aexp e and an integer value n for the variable x we
can evaluate e to finds its value eval.e; n/. It’s easy, and useful, to specify this
evaluation process with a recursive definition.
Definition 7.4.2. The evaluation function, eval W Aexp Z ! Z, is defined recursively on expressions e 2 Aexp as follows. Let n be any integer.
Base cases:
eval.x; n/ WWD n
(value of variable x is n),
(7.9)
eval.k; n/ WWD k
(value of numeral k is k, regardless of x.)
(7.10)
“mcs” — 2017/6/5 — 19:42 — page 223 — #231
7.4. Arithmetic Expressions
223
Constructor cases:
eval.[ e1 + e2 ] ; n/ WWD eval.e1 ; n/ C eval.e2 ; n/;
(7.11)
eval.[ e1 * e2 ] ; n/ WWD eval.e1 ; n/ eval.e2 ; n/;
(7.12)
eval.- [ e1 ] ; n/ WWD
eval.e1 ; n/:
(7.13)
For example, here’s how the recursive definition of eval would arrive at the value
of 3 C x 2 when x is 2:
eval.[ 3 + [ x * x ] ] ; 2/ D eval.3; 2/ C eval.[ x * x ] ; 2/ (by Def 7.4.2.7.11)
D 3 C eval.[ x * x ] ; 2/
(by Def 7.4.2.7.10)
D 3 C .eval.x; 2/ eval.x; 2//
(by Def 7.4.2.7.12)
D 3 C .2 2/
(by Def 7.4.2.7.9)
D 3 C 4 D 7:
Substituting into Aexp’s
Substituting expressions for variables is a standard operation used by compilers
and algebra systems. For example, the result of substituting the expression 3x for
x in the expression x.x 1/ would be 3x.3x 1/. We’ll use the general notation
subst.f; e/ for the result of substituting an Aexp f for each of the x’s in an Aexp
e. So as we just explained,
subst.3x; x.x
1// D 3x.3x
1/:
This substitution function has a simple recursive definition:
Definition 7.4.3. The substitution function from Aexp Aexp to Aexp is defined
recursively on expressions e 2 Aexp as follows. Let f be any Aexp.
Base cases:
subst.f; x/ WWD f
(subbing f for variable x just gives f ,)
(7.14)
subst.f; k/ WWD k
(subbing into a numeral does nothing.)
(7.15)
Constructor cases:
“mcs” — 2017/6/5 — 19:42 — page 224 — #232
224
Chapter 7
Recursive Data Types
subst.f; [ e1 + e2 ] / WWD [ subst.f; e1 / + subst.f; e2 /]
(7.16)
subst.f; [ e1 * e2 ] / WWD [ subst.f; e1 / * subst.f; e2 /]
(7.17)
subst.f; - [ e1 ] / WWD - [ subst.f; e1 /] :
(7.18)
Here’s how the recursive definition of the substitution function would find the
result of substituting 3x for x in the expression x.x 1/:
subst.3x; x.x
1//
D subst.[ 3 * x ] ; [ x * [ x + - [ 1] ] ] /
(unabbreviating)
D [ subst.[ 3 * x ] ; x/ *
subst.[ 3 * x ] ; [ x + - [ 1] ] /]
(by Def 7.4.3 7.17)
D [ [ 3 * x ] * subst.[ 3 * x ] ; [ x + - [ 1] ] /]
(by Def 7.4.3 7.14)
D [ [ 3 * x ] * [ subst.[ 3 * x ] ; x/
+ subst.[ 3 * x ] ; - [ 1] /] ]
(by Def 7.4.3 7.16)
D [ [ 3 * x ] * [ [ 3 * x ] + - [ subst.[ 3 * x ] ; 1/] ] ]
(by Def 7.4.3 7.14 & 7.18)
D [ [ 3 * x ] * [ [ 3 * x ] + - [ 1] ] ]
D 3x.3x
(by Def 7.4.3 7.15)
1/
(abbreviation)
Now suppose we have to find the value of subst.3x; x.x 1// when x D 2.
There are two approaches. First, we could actually do the substitution above to get
3x.3x 1/, and then we could evaluate 3x.3x 1/ when x D 2, that is, we could
recursively calculate eval.3x.3x 1/; 2/ to get the final value 30. This approach is
described by the expression
eval.subst.3x; x.x
1//; 2/:
(7.19)
In programming jargon, this would be called evaluation using the Substitution
Model. With this approach, the formula 3x appears twice after substitution, so
the multiplication 3 2 that computes its value gets performed twice.
The second approach is called evaluation using the Environment Model. Here, to
compute the value of (7.19), we evaluate 3x when x D 2 using just 1 multiplication
to get the value 6. Then we evaluate x.x 1/ when x has this value 6 to arrive at
the value 6 5 D 30. This approach is described by the expression
eval.x.x
1/; eval.3x; 2//:
(7.20)
The Environment Model only computes the value of 3x once, and so it requires one
fewer multiplication than the Substitution model to compute (7.20).
“mcs” — 2017/6/5 — 19:42 — page 225 — #233
7.4. Arithmetic Expressions
225
This is a good place to stop and work this example out yourself (Problem 7.26).
The fact that the final integer values of (7.19) and (7.20) agree is no surprise. The
substitution model and environment models will always produce the same final. We
can prove this by structural induction directly following the definitions of the two
approaches. More precisely, what we want to prove is
Theorem 7.4.4. For all expressions e; f 2 Aexp and n 2 Z,
eval.subst.f; e/; n/ D eval.e; eval.f; n//:
(7.21)
Proof. The proof is by structural induction on e.1
Base cases:
Case[x]
The left-hand side of equation (7.21) equals eval.f; n/ by this base case in
Definition 7.4.3 of the substitution function; the right-hand side also equals
eval.f; n/ by this base case in Definition 7.4.2 of eval.
Case[k].
The left-hand side of equation (7.21) equals k by this base case in Definitions 7.4.3 and 7.4.2 of the substitution and evaluation functions. Likewise,
the right-hand side equals k by two applications of this base case in the Definition 7.4.2 of eval.
Constructor cases:
Case[[ e1 + e2 ] ]
By the structural induction hypothesis (7.21), we may assume that for all
f 2 Aexp and n 2 Z,
eval.subst.f; ei /; n/ D eval.ei ; eval.f; n//
(7.22)
for i D 1; 2. We wish to prove that
eval.subst.f; [ e1 + e2 ] /; n/ D eval.[ e1 + e2 ] ; eval.f; n//:
(7.23)
The left-hand side of (7.23) equals
eval.[ subst.f; e1 / + subst.f; e2 /] ; n/
1 This
is an example of why it’s useful to notify the reader what the induction variable is—in this
case it isn’t n.
“mcs” — 2017/6/5 — 19:42 — page 226 — #234
226
Chapter 7
Recursive Data Types
by Definition 7.4.3.7.16 of substitution into a sum expression. But this equals
eval.subst.f; e1 /; n/ C eval.subst.f; e2 /; n/
by Definition 7.4.2.(7.11) of eval for a sum expression. By induction hypothesis (7.22), this in turn equals
eval.e1 ; eval.f; n// C eval.e2 ; eval.f; n//:
Finally, this last expression equals the right-hand side of (7.23) by Definition 7.4.2.(7.11) of eval for a sum expression. This proves (7.23) in this case.
Case[[ e1 * e2 ] ] Similar.
Case[ [ e1 ] ] Even easier.
This covers all the constructor cases, and so completes the proof by structural
induction.
7.5
Games as a Recursive Data Type
Chess, Checkers, Go, and Nim are examples of two-person games of perfect information. These are games where two players, Player-1 and Player-2, alternate
moves, and “perfect information” means that the situation at any point in the game
is completely visible to both players. In Chess, for example, the visible positions
of the pieces on the chess board completely determine how the rest of the game can
be played by each player. By contrast, most card games are not games of perfect
information because neither player can see the other’s hand.
In the section we’ll examine the win-lose two-person games of perfect information, WL-2PerGm. We will define WL-2PerGm as a recursive data type, and then
we will prove, by structural induction, a fundamental theorem about winning strategies for these games. The idea behind the recursive definition is to recognize that
the situation at any point during game play can itself be treated as the start of a new
game. This is clearest for the game of Nim.
A Nim game starts with several piles of stones. A move in the game consists of
removing some positive number of stones from a single pile. Player-1 and player-2
alternate making moves, and whoever takes the last stone wins. So if there is only
one pile, then the first player to move wins by taking the whole pile. On the other
hand, if the game starts with just two piles, each with the same number of stones,
“mcs” — 2017/6/5 — 19:42 — page 227 — #235
7.5. Games as a Recursive Data Type
227
then the player who moves second can guarantee a win simply by mimicking the
first player. For example, this means that if the first player removes three stones
from one pile, then the second player removes three stones from the other pile. At
this point, it’s worth thinking for a moment about why the mimicking strategy
guarantees a win for the second player.
We can think of the first move in a Nim game as simply picking another Nim
game with different piles of stone to play next. For the Nim game Nimh3;4;5i that
starts with piles of 3, 4 and 5 stones, the first player can remove between one and
three stones from the first pile leading to three possible piles of stones
h2; 4; 5i ; h1; 4; 5i ; h4; 5i :
Similarly, the first player has five possible ways to remove stones from the last pile,
leading to five possible piles of stones
h3; 4; 4i ; h3; 4; 3i ; h3; 4; 2i ; h3; 4; 1i ; h3; 4i :
So all the properties of Nimh3;4;5i are captured by the set of 3 C 4 C 5 D 12 Nim
games that can result from the first move.
With this idea in mind, we now give the formal definition.
Definition 7.5.1. The class WL-2PerGm of two-person win-lose games of perfect
information is defined recursively as follows:
Base case: win and lose are WL-2PerGm’s.
Constructor case: If G is a nonempty set of WL-2PerGm’s, then G is a WL-2PerGm
game. Each game M 2 G is called a possible first move of G.
A play of a WL-2PerGm game is a sequence of moves that ends with a win or
loss for the first player, or goes on forever without arriving at an outcome.2 More
formally:
Definition. A play of a WL-2PerGm game G and its outcome is defined recursively
on the definition of WL-2PerGm:
Base case: (G D win). The sequence hwini of length one is a play of G. Its
outcome is a win.
Base case: (G D lose). The sequence hlosei of length one is a play of G. Its
outcome is a loss.
2 In
English, “Nim game” might refer to the rules that define the game, but it might also refer
to a particular play of the game—as in the once famous third game in the 1961 movie Last Year at
Marienbad. It’s usually easy to figure out which way the phrase in being used, and we won’t worry
about it.
“mcs” — 2017/6/5 — 19:42 — page 228 — #236
228
Chapter 7
Recursive Data Types
Constructor case: (G is a nonempty set of WL-2PerGm’s). A play of G is a
sequence that starts with G followed by a play PM of some game M 2 G. The
outcome of the play, if any, is the outcome of PM .
The basic rules of some games do allow plays that go on forever. In Chess for
example, a player might just keep moving the same piece back and forth, and if his
opponent did the same, the play could go on forever.3 But the recursive definition
of WL-2PerGm games actually rules out the possibility of infinite play.
Lemma 7.5.2. Every play of a game G 2 WL-2PerGm has an outcome.
Proof. We prove Lemma 7.5.2 by structural induction, using the statement of the
Lemma as the induction hypothesis.
Base case: (G D win). There is only one play of G, namely the length one play
hwini, whose outcome is a win.
Base case: (G D lose). Likewise with the outcome being a loss.
Constructor case: (G is a nonempty set of WL-2PerGm’s). A play of G by definition consists G followed by a play PM for some M 2 G. By structural induction,
PM must be a sequence of some finite length n that ends with an outcome. So this
play of G is a length n C 1 sequence that finishes with the same outcome.
Among the games of Checker, Chess, Go and Nim, only Nim is genuinely a winlose game, The other games might end in a tie (draw, stalemate, jigo) rather than a
win or loss. However, by treating a tie in these games as a loss for the first player,
the results about win-lose games will apply to games with ties.
7.5.1
Game Strategies
A strategy for a player is a rule that tells the player which move to make whenever
it is their turn. More precisely, a strategy s is a function from games to games with
the property that s.G/ 2 G for all games G. A pair of strategies for the two players
determines exactly which moves the players choose, and so it determines a unique
play of the game, depending on who moves first.
A key question about a game is what strategy will ensure that a player will win.
The Player-1 wants a strategy whose outcome is guaranteed to be a win, and Player2 wants a strategy whose outcome is guaranteed to be a loss for Player-1.
3 Real chess tournaments rule this out by setting an advance limit on the number of moves, or by
forbidding repetitions of the same position more than twice.
“mcs” — 2017/6/5 — 19:42 — page 229 — #237
7.5. Games as a Recursive Data Type
7.5.2
229
Fundamental Theorem for Win-Lose Games
The Fundamental Theorem for WL-2PerGm games says that one of the players
always has a fixed “winning” strategy that guarantees a win against every possible
opponent strategy.
Thinking about Chess for instance, this seems surprising. Serious chess players
are typically secretive about their intended play strategies, believing that an opponent could take advantage of knowing their strategy. Their concern seems to be that
for any strategy they choose, their opponent coulda tailor a strategy to beat it.
But the Fundamental Theorem says otherwise. In theory, in any win-lose-tie
game like Chess or Checkers, each of the players will have a strategy that guarantees a win or a stalemate, even if the strategy is known to their opponent. That
is,
there is winning strategy for one of the players, or
both players have strategies that guarantee them at worst a draw.
Even though the Fundamental Theorem reveals a profound fact about games, it
has a very simple proof by structural induction.
Theorem 7.5.3. [Fundamental Theorem for Win-Lose Games] For any WL-2PerGm
game G, one of the players has a winning strategy.
Proof. The proof is by structural induction on the definition of a G 2 WL-2PerGm.
The induction hypothesis is that one of the players has a winning strategy for G.
Base case: (G D win or lose). Then there is only one possible strategy for each
player, namely, do nothing and finish with outcome G.
Constructor case: (G is a nonempty set of WL-2PerGm’s). By structural induction
we may assume that for each M 2 G one of the players has a winning strategy.
Notice that since players alternate moves, the first player in G becomes the second
player in M .
Now if there is a move M0 2 G where the second player in M0 has a winning
strategy, then the first player in G has a simple winning strategy: pick M0 as the
first move, and then follow the second player’s winning strategy for M0 .
On the other hand, if no M 2 G has a winning strategy for the second player in
M , then we can conclude by induction that every M 2 G has a winning strategy for
the first player in M . Now the second player in G has a simple winning strategy,
namely if the first player in G makes the move M , then the second player in G
should follow the follow the winning strategy for the first player in M .
“mcs” — 2017/6/5 — 19:42 — page 230 — #238
230
Chapter 7
Recursive Data Types
Infinite Games
So where do we come upon games with an infinite number of first moves? Well,
suppose we play a tournament of n chess games for some positive integer n. This
tournament will be a WL-2PerGm if we agree on a rule for combining the payoffs
of the n individual chess games into a final payoff for the whole tournament.
There still are only a finite number of possible moves at any stage of the n-game
chess tournament, but we can define a meta-chess-tournament, whose first move is
a choice of any positive integer n, after which we play an n-game tournament. Now
the meta-chess-tournament has an infinite number of first moves.
Of course only the first move in the meta-chess-tournament is infinite, but then
we could set up a tournament consisting of n meta-chess-tournaments. This would
be a game with n possible infinite moves. And then we could have a meta-metachess-tournament whose first move was to choose how many meta-chess-tournaments
to play. This meta-meta-chess-tournament will have an infinite number of infinite
moves. Then we could move on to meta-meta-meta-chess-tournaments . . . .
As silly or weird as these meta games may seem, their weirdness doesn’t disqualify the Fundamental Theorem: in each of these games, one of the players will
have winning strategy.
Notice that although Theorem 7.5.3 guarantees a winning strategy, its proof gives
no clue which player has it. For the Subset Takeaway Game of Problem 4.7 and
most familiar 2PerGm’s like Chess, Go, . . . , no one knows which player has a
winning strategy.4
7.6
Induction in Computer Science
Induction is a powerful and widely applicable proof technique, which is why we’ve
devoted two entire chapters to it. Strong induction and its special case of ordinary
induction are applicable to any kind of thing with nonnegative integer sizes—which
is an awful lot of things, including all step-by-step computational processes.
Structural induction then goes beyond number counting, and offers a simple,
natural approach to proving things about recursive data types and recursive computation.
In many cases, a nonnegative integer size can be defined for a recursively defined
datum, such as the length of a string, or the number of operations in an Aexp. It is
then possible to prove properties of data by ordinary induction on their size. But
4 Checkers
used to be in this list, but there has been a recent announcement that each player has a
strategy that forces a tie. (reference TBA)
“mcs” — 2017/6/5 — 19:42 — page 231 — #239
7.6. Induction in Computer Science
231
this approach often produces more cumbersome proofs than structural induction.
In fact, structural induction is theoretically more powerful than ordinary induction. However, it’s only more powerful when it comes to reasoning about infinite
data types—like infinite trees, for example—so this greater power doesn’t matter in
practice. What does matter is that for recursively defined data types, structural induction is a simple and natural approach. This makes it a technique every computer
scientist should embrace.
Problems for Section 7.1
Practice Problems
Problem 7.1.
The set OBT of Ordered Binary Trees is defined recursively as follows:
Base case: hleafi is an OBT, and
Constructor case: if R and S are OBT’s, then hnode; R; S i is an OBT.
If T is an OBT, let nT be the number of node labels in T and lT be the number
of leaf labels in T .
Prove by structural induction that for all T 2 OBT,
lT D nT C 1:
(7.24)
Class Problems
Problem 7.2.
Prove by structural induction on the recursive definition(7.1.1) of A that concatenation is associative:
.r s/ t D r .s t /
(7.25)
for all strings r; s; t 2 A .
Problem 7.3.
The reversal of a string is the string written backwards, for example, rev.abcde/ D
edcba.
(a) Give a simple recursive definition of rev.s/ based on the recursive definitions 7.1.1 of s 2 A and of the concatenation operation 7.1.3.
“mcs” — 2017/6/5 — 19:42 — page 232 — #240
232
Chapter 7
Recursive Data Types
(b) Prove that
rev.s t / D rev.t / rev.s/;
(7.26)
for all strings s; t 2 A . You may assume that concatenation is associative:
.r s/ t D r .s t /
for all strings r; s; t 2 A (Problem 7.2).
Problem 7.4.
The Elementary 18.01 Functions (F18’s) are the set of functions of one real variable
defined recursively as follows:
Base cases:
The identity function id.x/ WWD x is an F18,
any constant function is an F18,
the sine function is an F18,
Constructor cases:
If f; g are F18’s, then so are
1. f C g, fg, 2g ,
2. the inverse function f
1,
3. the composition f ı g.
(a) Prove that the function 1=x is an F18.
Warning: Don’t confuse 1=x D x 1 with the inverse id
id.x/. The inverse id 1 is equal to id.
1
of the identity function
(b) Prove by Structural Induction on this definition that the Elementary 18.01
Functions are closed under taking derivatives. That is, show that if f .x/ is an F18,
then so is f 0 WWD df =dx. (Just work out 2 or 3 of the most interesting constructor
cases; you may skip the less interesting ones.)
Problem 7.5.
Here is a simple recursive definition of the set E of even integers:
“mcs” — 2017/6/5 — 19:42 — page 233 — #241
7.6. Induction in Computer Science
233
Definition. Base case: 0 2 E.
Constructor cases: If n 2 E, then so are n C 2 and n.
Provide similar simple recursive definitions of the following sets:
(a) The set S WWD f2k 3m 5n 2 N j k; m; n 2 Ng.
(b) The set T WWD f2k 32kCm 5mCn 2 N j k; m; n 2 Ng.
(c) The set L WWD f.a; b/ 2 Z2 j .a b/ is a multiple of 3g.
Let L0 be the set defined by the recursive definition you gave for L in the previous
part. Now if you did it right, then L0 D L, but maybe you made a mistake. So let’s
check that you got the definition right.
(d) Prove by structural induction on your definition of L0 that
L0 L:
(e) Confirm that you got the definition right by proving that
L L0 :
(f) See if you can give an unambiguous recursive definition of L.
Problem 7.6.
Definition. The recursive data type binary-2PG of binary trees with leaf labels L
is defined recursively as follows:
Base case: hleaf; li 2 binary-2PG, for all labels l 2 L.
Constructor case: If G1 ; G2 2 binary-2PG, then
hbintree; G1 ; G2 i 2 binary-2PG:
The size jGj of G 2 binary-2PG is defined recursively on this definition by:
Base case:
j hleaf; li j WWD 1;
for all l 2 L:
Constructor case:
j hbintree; G1 ; G2 i j WWD jG1 j C jG2 j C 1:
“mcs” — 2017/6/5 — 19:42 — page 234 — #242
234
Chapter 7
Recursive Data Types
G
G1
win
G1,2
win
lose
win
Figure 7.1 A picture of a binary tree G.
For example, the size of the binary-2PG G pictured in Figure 7.1, is 7.
(a) Write out (using angle brackets and labels bintree, leaf, etc.) the binary-2PG
G pictured in Figure 7.1.
The value of flatten.G/ for G 2 binary-2PG is the sequence of labels in L of the
leaves of G. For example, for the binary-2PG G pictured in Figure 7.1,
flatten.G/ D .win; lose; win; win/:
(b) Give a recursive definition of flatten. (You may use the operation of concatenation (append) of two sequences.)
(c) Prove by structural induction on the definitions of flatten and size that
2 length.flatten.G// D jGj C 1:
(7.27)
Homework Problems
Problem 7.7.
The string reversal function, rev W A ! A has a simple recursive definition.
Base case: rev./ WWD .
Constructor case: rev.as/ WWD rev.s/a for s 2 A and a 2 A.
“mcs” — 2017/6/5 — 19:42 — page 235 — #243
7.6. Induction in Computer Science
235
A string s is a palindrome when rev.s/ D s. The palindromes also have a simple
recursive definition as the set RecPal.
Base cases: 2 RecPal and a 2 RecPal for a 2 A.
Constructor case: If s 2 RecPal, then asa 2 RecPal for a 2 A.
Verifying that the two definitions agree offers a nice exercise in structural induction and also induction on length of strings. The verification rests on three basic
properties of concatenation and reversal proved in separate problems 7.2 and 7.3.
Fact.
.rs D uv AND jrj D juj/ IFF .r D u AND s D v/
r .s t / D .r s/ t
rev.st / D rev.t / rev.s/
(7.28)
(7.29)
(7.30)
(a) Prove that s D rev.s/ for all s 2 RecPal.
(b) Prove conversely that if s D rev.s/, then s 2 RecPal.
Hint: By induction on n D jsj.
Problem 7.8.
Let m; n be integers, not both zero. Define a set of integers, Lm;n , recursively as
follows:
Base cases: m; n 2 Lm;n .
Constructor cases: If j; k 2 Lm;n , then
1.
j 2 Lm;n ,
2. j C k 2 Lm;n .
Let L be an abbreviation for Lm;n in the rest of this problem.
(a) Prove by structural induction that every common divisor of m and n also divides every member of L.
(b) Prove that any integer multiple of an element of L is also in L.
(c) Show that if j; k 2 L and k ¤ 0, then rem.j; k/ 2 L.
(d) Show that there is a positive integer g 2 L that divides every member of L.
Hint: The least positive integer in L.
“mcs” — 2017/6/5 — 19:42 — page 236 — #244
236
Chapter 7
Recursive Data Types
Figure 7.2 Constructing the Koch Snowflake.
(e) Conclude that g from part (d) is gcd.m; n/, the greatest common divisor, of m
and n.
Problem 7.9.
Definition. Define the number #c .s/ of occurrences of the character c 2 A in the
string s recursively on the definition of s 2 A :
base case: #c ./ WWD 0.
constructor case:
(
#c .s/
if a ¤ c;
#c .ha; si/ WWD
1 C #c .s/ if a D c:
Prove by structural induction that for all s; t 2 A and c 2 A
#c .s t / D #c .s/ C #c .t /:
Problem 7.10.
Fractals are an example of mathematical objects that can be defined recursively.
In this problem, we consider the Koch snowflake. Any Koch snowflake can be
constructed by the following recursive definition.
Base case: An equilateral triangle with a positive integer side length is a
Koch snowflake.
Constructor case: Let K be a Koch snowflake, and let l be a line segment
on the snowflake. Remove the middle third of l, and replace it with two line
segments of the same length jlj, as is done in Figure 7.2
The resulting figure is also a Koch snowflake.
Prove
pby structural induction that the area inside any Koch snowflake is of the
form q 3, where q is a rational number.
“mcs” — 2017/6/5 — 19:42 — page 237 — #245
7.6. Induction in Computer Science
237
Problem 7.11.
The set RBT of Red-Black Trees is defined recursively as follows:
Base cases:
hredi 2 RBT, and
hblacki 2 RBT.
Constructor cases: A; B are RBT’s, then
if A; B start with black, then hred; A; Bi is an RBT.
if A; B start with red, then hblack; A; Bi is an RBT.
For any RBT T , let
rT be the number of red labels in T ,
bT be the number of black labels in T , and
nT WWD rT C bT be the total number of labels in T .
Prove that
If T starts with a red label; then
2nT C 1
nT
rT ;
3
3
(7.31)
Hint:
n=3 r
IFF
.2=3/n n
r
Exam Problems
Problem 7.12.
The Arithmetic Trig Functions (Atrig’s) are the set of functions of one real variable
defined recursively as follows:
Base cases:
The identity function id.x/ WWD x is an Atrig,
any constant function is an Atrig,
the sine function is an Atrig,
“mcs” — 2017/6/5 — 19:42 — page 238 — #246
238
Chapter 7
Recursive Data Types
Constructor cases:
If f; g are Atrig’s, then so are
1. f C g
2. f g
3. the composition f ı g.
Prove by structural induction on this definition that if f .x/ is an Atrig, then so is
f 0 WWD df =dx.
Problem 7.13.
Definition. The set RAF of rational functions of one real variable is the set of
functions defined recursively as follows:
Base cases:
The identity function, id.r/ WWD r for r 2 R (the real numbers), is an RAF,
any constant function on R is an RAF.
Constructor cases: If f; g are RAF’s, then so is f ~ g, where ~ is one of the
operations
1. addition C,
2. multiplication or
3. division =.
(a) Describe how to construct functions e; f; g 2 RAF such that
e ı .f C g/ ¤ .e ı f / C .e ı g/:
(7.32)
(b) Prove that for all real-valued functions e; f; g (not just those in RAF):
.e ~ f / ı g D .e ı g/ ~ .f ı g/;
Hint: .e ~ f /.x/ WWD e.x/ ~ f .x/.
(7.33)
“mcs” — 2017/6/5 — 19:42 — page 239 — #247
7.6. Induction in Computer Science
239
(c) Let predicate P .h/ be the following predicate on functions h 2 RAF:
P .h/ WWD 8g 2 RAF: h ı g 2 RAF:
Prove by structural induction on the definition of RAF that P .h/ holds for all h 2
RAF.
Make sure to indicate explicitly
each of the base cases, and
each of the constructor cases.
Problem 7.14.
The 2-3-averaged numbers are a subset, N23, of the real interval Œ0; 1 defined
recursively as follows:
Base cases: 0; 1 2 N23.
Constructor case: If a; b are in N23, then so is L.a; b/ where
L.a; b/ WWD
2a C 3b
:
5
(a) Use ordinary induction or the Well-Ordering Principle to prove that
n
3
2 N23
5
for all nonnegative integers n.
(b) Prove by Structural Induction that the product of two 2-3-averaged numbers is
also a 2-3-averaged number.
Hint: Prove by structural induction on c that, if d 2 N23, then cd 2 N23.
Problem 7.15.
This problem is about binary strings s 2 f0; 1g .
Let’s call a recursive definition of a set of strings cat-OK when all its constructors
are defined as concatenations of strings.5
5 The
concatenation of two strings x and y, written xy, is the string obtained by appending x to
the left end of y. For example, the concatenation of 01 and 101 is 01101.
“mcs” — 2017/6/5 — 19:42 — page 240 — #248
240
Chapter 7
Recursive Data Types
For example, the set, One1, of strings with exactly one 1 has the cat-OK definition:
Base case: The length-one string 1 is in One1.
Constructor case: If s is in One1, then so is 0s and s0.
(a) Give a cat-OK definition of the set E of even length strings consisting solely
of 0’s.
(b) Let rev.s/ be the reversal of the string s. For example, rev.001/ D 100. A
palindrome is a string s such that s D rev.s/. For example, 11011 and 010010
are palindromes.
Give a cat-OK definition of the palindromes.
(c) Give a cat-OK definition of the set P of strings consisting solely of 0’s whose
length is a power of two.
Problems for Section 7.2
Practice Problems
Problem 7.16.
Define the sets F1 and F2 recursively:
F1 :
– 5 2 F1 ,
– if n 2 F1 , then 5n 2 F1 .
F2 :
– 5 2 F2 ,
– if n; m 2 F1 , then nm 2 F2 .
(a) Show that one of these definitions is technically ambiguous. (Remember that
“ambiguous recursive definition” has a technical mathematical meaning which does
not imply that the ambiguous definition is unclear.)
(b) Briefly explain what advantage unambiguous recursive definitions have over
ambiguous ones.
“mcs” — 2017/6/5 — 19:42 — page 241 — #249
7.6. Induction in Computer Science
241
(c) A way to prove that F1 D F2 , is to show first that F1 F2 and second that
F2 F1 . One of these containments follows easily by structural induction. Which
one? What would be the induction hypothesis? (You do not need to complete a
proof.)
Problem 7.17. (a) To prove that the set RecMatch, of matched strings of Definition 7.2.1 equals the set AmbRecMatch of ambiguous matched strings of Definition 7.2.4, you could first prove that
8r 2 RecMatch: r 2 AmbRecMatch;
and then prove that
8u 2 AmbRecMatch: u 2 RecMatch:
Of these two statements, indicate the one that would be simpler to prove by structural induction directly from the definitions.
(b) Suppose structural induction was being used to prove that AmbRecMatch RecMatch. Indicate the one predicate below that would fit the format for a structural
induction hypothesis in such a proof.
P0 .n/ WWD jsj n IMPLIES s 2 RecMatch.
P1 .n/ WWD jsj n IMPLIES s 2 AmbRecMatch.
P2 .s/ WWD s 2 RecMatch.
P3 .s/ WWD s 2 AmbRecMatch.
P4 .s/ WWD .s 2 RecMatch IMPLIES s 2 AmbRecMatch/.
(c) The recursive definition AmbRecMatch is ambiguous because it allows the
s t constructor to apply when s or t is the empty string. But even fixing that,
ambiguity remains. Demonstrate this by giving two different derivations for the
string ”[ ] [ ] [ ] according to AmbRecMatch but only using the s t constructor
when s ¤ and t ¤ .
Class Problems
Problem 7.18.
Let p be the string [ ] . A string of brackets is said to be erasable iff it can be
“mcs” — 2017/6/5 — 19:42 — page 242 — #250
242
Chapter 7
Recursive Data Types
reduced to the empty string by repeatedly erasing occurrences of p. For example,
to erase the string
[[[]][]][];
start by erasing the three occurrences of p to obtain
[[]]:
Then erase the single occurrence of p to obtain,
[];
which can now be erased to obtain the empty string .
On the other hand the string
[]][[[[[]]
(7.34)
is not erasable, because when we try to erase, we get stuck. Namely, start by erasing
the two occurrences of p in (7.34) to obtain
][[[[]:
The erase the one remaining occurrence of p to obtain.
][[[:
At this point we are stuck with no remaining occurrences of p. 6
Let Erasable be the set of erasable strings of brackets. Let RecMatch be the
recursive data type of strings of matched brackets given in Definition 7.2.1
(a) Use structural induction to prove that
RecMatch Erasable:
(b) Supply the missing parts (labeled by “(*)”) of the following proof that
Erasable RecMatch:
6 Notice
that there are many ways to erase a string, depending on when and which occurrences
of p are chosen to be erased. It turns out that given any initial string, the final string reached after
performing all possible erasures will be the same, no matter how erasures are performed. We take
this for granted here, although it is not altogether obvious. (See Problem 6.28 for a proof).
“mcs” — 2017/6/5 — 19:42 — page 243 — #251
7.6. Induction in Computer Science
243
Proof. We prove by strong induction that every length n string in Erasable is also
in RecMatch. The induction hypothesis is
P .n/ WWD 8x 2 Erasable: jxj D n IMPLIES x 2 RecMatch:
Base case:
(*) What is the base case? Prove that P is true in this case.
Inductive step: To prove P .n C 1/, suppose jxj D n C 1 and x 2 Erasable. We
need to show that x 2 RecMatch.
Let’s say that a string y is an erase of a string z iff y is the result of erasing a single
occurrence of p in z.
Since x 2 Erasable and has positive length, there must be an erase, y 2 Erasable,
of x. So jyj D n 1 0, and since y 2 Erasable, we may assume by induction
hypothesis that y 2 RecMatch.
Now we argue by cases:
Case (y is the empty string):
(*) Prove that x 2 RecMatch in this case.
Case (y D [ s ] t for some strings s; t 2 RecMatch): Now we argue by subcases.
Subcase(x D py):
(*) Prove that x 2 RecMatch in this subcase.
Subcase (x is of the form [ s 0 ] t where s is an erase of s 0 ):
Since s 2 RecMatch, it is erasable by part (b), which implies that s 0 2
Erasable. But js 0 j < jxj, so by induction hypothesis, we may assume that
s 0 2 RecMatch. This shows that x is the result of the constructor step of
RecMatch, and therefore x 2 RecMatch.
Subcase (x is of the form [ s ] t 0 where t is an erase of t 0 ):
(*) Prove that x 2 RecMatch in this subcase.
(*) Explain why the above cases are sufficient.
This completes the proof by strong induction on n, so we conclude that P .n/ holds
for all n 2 N. Therefore x 2 RecMatch for every string x 2 Erasable. That is,
Erasable RecMatch. Combined with part (a), we conclude that
Erasable D RecMatch:
“mcs” — 2017/6/5 — 19:42 — page 244 — #252
244
Chapter 7
Recursive Data Types
Problem 7.19. (a) Prove that the set RecMatch of matched strings of Definition 7.2.1
is closed under string concatenation. Namely, 8s; t 2 RecMatch: s t 2 RecMatch.
(b) Prove AmbRecMatch RecMatch, where AmbRecMatch is the set of ambiguous matched strings of Definition 7.2.4.
(c) Prove that RecMatch D AmbRecMatch.
Homework Problems
Problem 7.20.
One way to determine if a string has matching brackets, that is, if it is in the set,
RecMatch, of Definition 7.2.1 is to start with 0 and read the string from left to right,
adding 1 to the count for each left bracket and subtracting 1 from the count for each
right bracket. For example, here are the counts for two sample strings:
[
]
0 1 0
[
[
0 1 2
]
[
[
[
[
[
]
]
]
]
1 0 1 2 3 4 3 2 1 0
[
]
]
[
]
]
[
]
3 2 1 2 1 0 1 0
A string has a good count if its running count never goes negative and ends with 0.
So the second string above has a good count, but the first one does not because its
count went negative at the third step. Let
GoodCount WWD fs 2 f] ; [ g j s has a good countg:
The empty string has a length 0 running count we’ll take as a good count by
convention, that is, 2 GoodCount. The matched strings can now be characterized
precisely as this set of strings with good counts.
(a) Prove that GoodCount contains RecMatch by structural induction on the definition of RecMatch.
(b) Conversely, prove that RecMatch contains GoodCount.
Hint: By induction on the length of strings in GoodCount. Consider when the
running count equals 0 for the second time.
Problem 7.21.
Divided Equilateral Triangles (DETs) were defined in Problem 5.10 as follows:
“mcs” — 2017/6/5 — 19:42 — page 245 — #253
7.6. Induction in Computer Science
245
[h]
Figure 7.3 DET T 0 from Four Copies of DET T
[h]
Figure 7.4 Trapezoid from Three Triangles
Base case: A single equilateral triangle is a DET whose only subtriangle is
itself.
If T WWD
is a DET, then the equilateral triangle T 0 built out of four
copies of T as shown in in Figure 7.3 is also a DET, and the subtriangles of
T 0 are exactly the subtriangles of each of the copies of T .
Properties of DETs were proved earlier by induction on the length of a side of the
triangle. Recognizing that the definition of DETs is recursive, we can instead prove
properties of DETs by structural induction.
(a) Prove by structural induction that a DET with one of its corner subtriangles
removed can be tiled with trapezoids built out of three subtriangles as in Figure 7.4.
(b) Explain why a DET with a triangle removed from the middle of one side can
also be tiled by trapezoids.
(c) In tiling a large square using L-shaped blocks as described in Section 5.1.5,
there was a tiling with any single subsquare removed. Part (b) indicates that trapezoidtilings are possible for DETs with a non-corner subtriangle removed, so it’s natural
to make the mistaken guess that DETs have a corresponding property:
False Claim. A DET with any single subtriangle removed can be trapezoid-tiled.
We can try to prove the claim by structural induction as in part (a).
“mcs” — 2017/6/5 — 19:42 — page 246 — #254
246
Chapter 7
Recursive Data Types
Bogus proof. The claim holds vacuously in the base case of a DET with a single
subtriangle.
Now let T 0 be a DET made of four copies of a DET T , and suppose we remove an
arbitrary subtriangle from T 0 .
The removed subtriangle must be a subtriangle of one of the copies of T . The
copies are the same, so for definiteness we assume the subtriangle was removed
from copy 1. Then by structural induction hypothesis, copy 1 can be trapezoidtiled, and then the other three copies of T can be trapezoid-tiled exactly as in the
solution to part(a). This yields a complete trapezoid-tiling of T 0 with the arbitrary
subtriangle removed.
We conclude by structural induction that any DET with any subtriangle removed
can be trapezoid-tiled.
What’s wrong with the proof?
Hint: Find a counter-example and show where the proof breaks down.
We don’t know if there is a simple characterization of exactly which subtriangles
can be removed to allow a trapezoid tiling.
Problem 7.22.
A binary word is a finite sequence of 0’s and 1’s. In this problem, we’ll simply call
them “words.” For example, .1; 1; 0/ and .1/ are words of length three and one,
respectively. We usually omit the parentheses and commas in the descriptions of
words, so the preceding binary words would just be written as 110 and 1.
The basic operation of placing one word immediately after another is called concatentation. For example, the concatentation of 110 and 1 is 1101, and the concatentation of 110 with itself is 110110.
We can extend this basic operation on words to an operation on sets of words. To
emphasize the distinction between a word and a set of words, from now on we’ll
refer to a set of words as a language. Now if R and S are languages, then R S is
the language consisting of all the words you can get by concatenating a word from
R with a word from S . That is,
R S WWD frs j r 2 R AND s 2 S g:
For example,
f0; 00g f00; 000g D f000; 0000; 00000g
“mcs” — 2017/6/5 — 19:42 — page 247 — #255
7.6. Induction in Computer Science
247
Another example is D D, abbreviated as D 2 , where D WWD f1; 0g.
D 2 D f00; 01; 10; 11g:
In other words, D 2 is the language consisting of all the length-two words. More
generally, D n will be the language of length-n words.
If S is a language, the language you can get by concatenating any number of
copies of words in S is called S —pronounced “S star.” (By convention, the empty
word always included in S .) For example, f0; 11g is the language consisting
of all the words you can make by stringing together 0’s and 11’s. This language
could also be described as consisting of the words whose blocks of 1’s are always
of even length. Another example is .D 2 / , which consists of all the even length
words. Finally, the language B of all binary words is just D .
The Concatenation-Definable (C-D) languages are defined recursively:
Base case: Every finite language is a C-D.
Constructor cases: If L and M are C-D’s, then
L M;
L [ M;
and L
are C-D’s.
Note that the -operation is not allowed. For this reason, the C-D languages are
also called the “star-free languages,” [36].
Lots of interesting languages turn out to be concatenation-definable, but some
very simple languages are not. This problem ends with the conclusion that the
language f00g of even length words whose bits are all 0’s is not a C-D language.
(a) Show that the set B of all binary words is C-D. Hint: The empty set is finite.
Now a more interesting example of a C-D set is the language of all binary words
that include three consecutive 1’s:
B111B:
Notice that the proper expression here is “B f111gB.” But it causes no confusion
and helps readability to omit the dots in concatenations and the curly braces for sets
with only one element.
(b) Show that the language consisting of the binary words that start with 0 and
end with 1 is C-D.
(c) Show that 0 is C-D.
“mcs” — 2017/6/5 — 19:42 — page 248 — #256
248
Chapter 7
Recursive Data Types
(d) Show that if R and S are C-D, then so is R \ S .
(e) Show that f01g is C-D.
Let’s say a language S is 0-finite when it includes only a finite number of words
whose bits are all 0’s, that is, when S \ 0 is a finite set of words. A langauge S is
0-boring—boring, for short—when either S or S is 0-finite.
(f) Explain why f00g is not boring.
(g) Verify that if R and S are boring, then so is R [ S .
(h) Verify that if R and S are boring, then so is R S .
Hint: By cases: whether R and S are both 0-finite, whether R or S contains no
all-0 words at all (including the empty word ), and whether neither of these cases
hold.
(i) Conclude by structural induction that all C-D languages are boring.
So we have proved that the set .00/ of even length all-0 words is not a C-D
language.
Problem 7.23.
We can explain in a simple and precise way how digital circuits work, and gain the
powerful proof method of structural induction to verify their properties, by defining
digital circuits as a recursive data type DigCirc. The definition is a little easier to
state if all the gates in the circuit take two inputs, so we will use the two-input NOR
gate rather than a one-input NOT, and let the set of gates be
Gates WWD fNOR; AND; OR; XORg:
A digital circuit will be a recursively defined list of gate connections of the form
.x; y; G; I / where G is a gate, x and y are the input wires, and I is the set of wires
that the gate output feeds into as illustrated in Figure 7.5.
Formally, we let W be a set w0 ; w1 ; : : : whose elements are called wires, and
O … W be an object called the output.
Definition. The set of digital circuit DigCirc, and their inputs and internal wires,
are defined recursively as follows:
“mcs” — 2017/6/5 — 19:42 — page 249 — #257
7.6. Induction in Computer Science
Figure 7.5
249
Digital Circuit Constructor Step
Base case: If x; y 2 W , then C 2 DigCirc, where
C D list..x; y; G; fOg// for some G 2 Gates;
inputs.C/ WWD fx; yg;
internal.C/ WWD ;:
Constructor cases: If
C 2 DigCirc;
I inputs.C/; I ¤ ;;
x; y 2 W
.I [ internal.C//
then D 2 DigCirc, where
D D cons..x; y; G; I/; C/ for some G 2 Gates;
inputs.D/ WWD fx; yg [ .inputs.C/
I /;
internal.D/ WWD internal.C/ [ I:
For any circuit C define
wires.C/ WWD inputs.C/ [ internal.C/ [ fOg:
“mcs” — 2017/6/5 — 19:42 — page 250 — #258
250
Chapter 7
Recursive Data Types
A wire assignment for C is a function
˛ W wires.C/ ! fT; Fg
such that for each gate connection .x; y; G; I / 2 C ,
˛.i / D .˛.x/ G ˛.y// for all i 2 I:
(a) Define an environment for C to be a function e W inputs.C/ ! fT; Fg. Prove
that if two wire assignments for C are equal for each wire in inputs.C/, then the
wire assignments are equal for all wires.
Part (a) implies that for any environment e for C , there is a unique wire assignment ˛e such that
˛e .w/ D e.w/ for all w 2 inputs.C/:
So for any input environment e, the circuit computes a unique output
eval.C; e/ WWD ˛e .O/:
Now suppose F is a propositional formula whose propositional variables are the
input wires of some circuit C . Then C and F are defined to be equivalent iff
eval.C; e/ D eval.F; e/
for all environments e for C .
(b) Define a function E.C / recursively on the definition of circuit C , such that
E.C / is a propositional formula equivalent to C . Then verify the recursive definition by proving the equivalence using structural induction.
(c) Give examples where E.C / is exponentially larger than C .
Exam Problems
Problem 7.24.
Let P be a propositional variable.
(a) Show how to express NOT.P / using P and a selection from among the constant True, and the connectives XOR and AND.
The use of the constant True above is essential. To prove this, we begin with
a recursive definition of XOR-AND formulas that do not use True, called the PXA
formulas.
“mcs” — 2017/6/5 — 19:42 — page 251 — #259
7.6. Induction in Computer Science
251
Definition. Base case: The propositional variable P is a PXA formula.
Constructor cases If R; S 2 PXA, then
R XOR S ,
R AND S
are PXA’s.
For example,
...P XOR P / AND P / XOR .P AND P // XOR .P XOR P /
is a PXA.
(b) Prove by structural induction on the definition of PXA that every PXA formula
A is equivalent to P or to False.
Problems for Section 7.3
Homework Problems
Problem 7.25.
One version of the the Ackermann function A W N2 ! N is defined recursively by
the following rules:
A.m; n/ WWD 2n
A.m; n/ WWD A.m
1; A.m; n
if m D 0 or n 1,
(A-base)
otherwise:
(AA)
1//
Prove that if B W N2 ! N is a partial function that satisfies this same definition,
then B is total and B D A.
Problems for Section 7.4
Practice Problems
Problem 7.26. (a) Write out the evaluation of
eval.subst.3x; x.x
1//; 2/
according to the Environment Model and the Substitution Model, indicating where
the rule for each case of the recursive definitions of eval.; / and ŒWD] or substitution
is first used. Compare the number of arithmetic operations and variable lookups.
“mcs” — 2017/6/5 — 19:42 — page 252 — #260
252
Chapter 7
Recursive Data Types
(b) Describe an example along the lines of part (a) where the Environment Model
would perform 6 fewer multiplications than the Substitution model. You need not
carry out the evaluations.
(c) Describe an example along the lines of part (a) where the Substitution Model
would perform 6 fewer multiplications than the Environment model. You need not
carry out the evaluations.
Class Problems
Problem 7.27.
In this problem we’ll need to be careful about the propositional operations that
apply to truth values and the corresponding symbols that appear in formulas. We’ll
restrict ourselves to formulas with symbols And and Not that correspond to the
operations AND, NOT. We will also allow the constant symbols True and False.
(a) Give a simple recursive definition of propositional formula F and the set
pvar.F / of propositional variables that appear in it.
Let V be a set of propositional variables. A truth environment e over V assigns
truth values to all these variables. In other words, e is a total function,
e W V ! fT; Fg:
(b) Give a recursive definition of the truth value, eval.F; e/, of propositional formula F in an environment e over a set of variables V pvar.F /.
Clearly the truth value of a propositional formula only depends on the truth values of the variables in it. How could it be otherwise? But it’s good practice to work
out a rigorous definition and proof of this assumption.
(c) Give an example of a propositional formula containing the variable P but
whose truth value does not depend on P . Now give a rigorous definition of the assertion that “the truth value of propositional formula F does not depend on propositional variable P .”
Hint: Let e1 ; e2 be two environments whose values agree on all variables other than
P.
(d) Give a rigorous definition of the assertion that “the truth value of a propositional formula only depends on the truth values of the variables that appear in it,”
and then prove it by structural induction on the definition of propositional formula.
(e) Now we can formally define F being valid. Namely, F is valid iff
8e: eval.F; e/ D T:
“mcs” — 2017/6/5 — 19:42 — page 253 — #261
7.6. Induction in Computer Science
253
Give a similar formal definition of formula G being unsatisfiable. Then use the
definition of eval to prove that a formula F is valid iff Not.F / is unsatisfiable.
Exam Problems
Problem 7.28.
A class of propositional formulas called the Multivariable AND-OR (MVAO) formulas are defined recursively as follows:
Definition. Base cases: A single propositional variable, and the constants True
and False are MVAO formulas.
Constructor cases: If G; H 2 MVAO, then .G ANDH / and .G ORH / are MVAO’s.
For example,
...P OR Q/ AND P / OR .R AND True// OR .Q OR False/
is a MVAO.
Definition. A propositional formula G is False-decreasing when substituting the
constant False for some occurrences of its variables makes the formula “more
false.” More precisely, if G f is the result of replacing some occurrences of variables in G by False, then any truth assignment that makes G false also makes G f
false.
For example, the formula consisting of a single variable P is False-decreasing
since P f is the formula False. The formula G WWD P is not False-decreasing since
G f is the formula False which is true even under a truth assignment where G is
false.
Prove by structural induction that every MVAO formula F is False-decreasing.
Homework Problems
Problem 7.29. (a) Give a recursive definition of a function erase.e/ that erases all
the symbols in e 2 Aexp but the brackets. For example
erase.[ [ 3 * [ x * x ] ] + [ [ 2 * x ] + 1] ] / D [ [ [ ] ] [ [ 2 * x ] + 1] ] :
(b) Prove that erase.e/ 2 RecMatch for all e 2 Aexp.
(c) Give an example of a small string s 2 RecMatch such that [ s ] ¤ erase.e/ for
any e 2 Aexp.
“mcs” — 2017/6/5 — 19:42 — page 254 — #262
254
Chapter 7
Recursive Data Types
Problems for Section 7.5
Practice Problems
Problem 7.30.
In the game tree for the game Tic-Tac-Toe, the root has nine children corresponding
to the nine boxes that the first player could mark with an “X”.
Each of these nine nodes will have eight children in the second level of the tree,
indicating where the second player can mark his “O”, giving a total of 72 nodes.
Answer the following questions about the game tree for Tic-Tac-Toe.
(a) How many nodes will be in the third level of the tree?
(b) What is the first level where this simple pattern of calculating nodes stops
working?
Homework Problems
Problem 7.31.
We’re going to characterize a large category of games as a recursive data type and
then prove, by structural induction, a fundamental theorem about game strategies.
We are interested in two person games of perfect information that end with a numerical score. Chess and Checkers would count as value games using the values
1; 1; 0 for a win, loss or draw for the first player. The game of Go really does end
with a score based on the number of white and black stones that remain at the end.
Here’s the formal definition:
Definition. Let V be a nonempty set of real numbers. The class VG of V -valued
two-person deterministic games of perfect information is defined recursively as follows:
Base case: A value v 2 V is a VG known as a payoff.
Constructor case: If G is a nonempty set of VG’s, then G is a VG. Each game
M 2 G is called a possible first move of G.
A strategy for a player is a rule that tells the player which move to make whenever it is their turn. That is, a strategy is a function s from games to games with the
property that s.G/ 2 G for all games G. Given which player has the first move, a
pair of strategies for the two players determines exactly which moves the players
will choose. So the strategies determine a unique play of the game and a unique
“mcs” — 2017/6/5 — 19:42 — page 255 — #263
7.6. Induction in Computer Science
255
payoff.7
The max-player wants a strategy that guarantees as high a payoff as possible,
and the min-player wants a strategy that guarantees as low a payoff as possible.
The Fundamental Theorem for deterministic games of perfect information says
that in any game, each player has an optimal strategy, and these strategies lead to
the same payoff. More precisely,
Theorem (Fundamental Theorem for VG’s). Let V be a finite set of real numbers
and G be a V -valued VG. Then there is a value v 2 V , called a max-value maxG
for G, such that if the max-player moves first,
the max-player has a strategy that will finish with a payoff of at least maxG ,
no matter what strategy the min-player uses, and
the min-player has a strategy that will finish with a payoff of at most maxG ,
no matter what strategy the max-player uses.
It’s worth a moment for the reader to observe that the definition of maxG implies
that if there is one for G, it is unique. So if the max-player has the first move,
the Fundamental Theorem means that there’s no point in playing the game: the
min-player may just as well pay the max-value to the max-player.
(a) Prove the Fundamental Theorem for VG’s.
Hint: VG’s are a recursively defined data type, so the basic method for proving that
all VG’s have some property is structural induction on the definition of VG. Since
the min-player moves first in whichever game the max-player picks for their first
move, the induction hypothesis will need to cover that case as well.
(b) (OPTIONAL). State some reasonable generalization of the Fundamental Theorem to games with an infinite set V of possible payoffs.
Problem 7.32.
Nim is a two-person game that starts with some piles of stones. A player’s move
consists of removing one or more stones from a single pile. The players alternate
making moves, and whoever takes the last stone wins.
It turns out there is a winning strategy for one of the players that is easy to carry
out but is not so obvious.
7 We
take for granted the fact that no VG has an infinite play. The proof of this by structural
induction is essentially the same as that for win-lose games given in Lemma 7.5.2.
“mcs” — 2017/6/5 — 19:42 — page 256 — #264
256
Chapter 7
Recursive Data Types
To explain the winning strategy, we need to think of a number in two ways: as
a nonnegative integer and as the bit string equal to the binary representation of the
number—possibly with leading zeroes.
For example, the XOR of numbers r; s; ::: is defined in terms of their binary representations: combine the corresponding bits of the binary representations of r; s; :::
using XOR, and then interpret the resulting bit-string as a number. For example,
2 XOR 7 XOR 9 D 12
because, taking XOR’s down the columns, we have
0
0
1
1
0
1
0
1
1
1
0
0
0
1
1
0
(binary rep of 2)
(binary rep of 7)
(binary rep of 9)
(binary rep of 12)
This is the same as doing binary addition of the numbers, but throwing away the
carries (see Problem 3.6).
The XOR of the numbers of stones in the piles is called their Nim sum. In this
problem we will verify that if the Nim sum is not zero on a player’s turn, then the
player has a winning strategy. For example, if the game starts with five piles of
equal size, then the first player has a winning strategy, but if the game starts with
four equal-size piles, then the second player can force a win.
(a) Prove that if the Nim sum of the piles is zero, then any one move will leave a
nonzero Nim sum.
(b) Prove that if there is a pile with more stones than the Nim sum of all the other
piles, then there is a move that makes the Nim sum equal to zero.
(c) Prove that if the Nim sum is not zero, then one of the piles is bigger than the
Nim sum of the all the other piles.
Hint: Notice that the largest pile may not be the one that is bigger than the Nim
sum of the others; three piles of sizes 2,2,1 is an example.
(d) Conclude that if the game begins with a nonzero Nim sum, then the first player
has a winning strategy.
Hint: Describe a preserved invariant that the first player can maintain.
(e) (Extra credit) Nim is sometimes played with winners and losers reversed, that
is, the person who takes the last stone loses. This is called the misère version of the
game. Use ideas from the winning strategy above for regular play to find one for
misère play.
“mcs” — 2017/6/5 — 19:42 — page 257 — #265
8
Infinite Sets
This chapter is about infinite sets and some challenges in proving things about
them.
Wait a minute! Why bring up infinity in a Mathematics for Computer Science
text? After all, any data set in a computer is limited by the size of the computer’s
memory, and there is a bound on the possible size of computer memory, for the
simple reason that the universe is (or at least appears to be) bounded. So why not
stick with finite sets of some large, but bounded, size? This is a good question, but
let’s see if we can persuade you that dealing with infinite sets is inevitable.
You may not have noticed, but up to now you’ve already accepted the routine use
of the integers, the rationals and irrationals, and sequences of them. These are all
infinite sets. Further, do you really want Physics or the other sciences to give up the
real numbers on the grounds that only a bounded number of bounded measurements
can be made in a bounded universe? It’s pretty convincing—and a lot simpler—to
ignore such big and uncertain bounds (the universe seems to be getting bigger all
the time) and accept theories using real numbers.
Likewise in computer science, it’s implausible to think that writing a program to
add nonnegative integers with up to as many digits as, say, the stars in the sky—
billions of galaxies each with billions of stars—would be different from writing a
program that would add any two integers, no matter how many digits they had. The
same is true in designing a compiler: it’s neither useful nor sensible to make use of
the fact that in a bounded universe, only a bounded number of programs will ever
be compiled.
Infinite sets also provide a nice setting to practice proof methods, because it’s
harder to sneak in unjustified steps under the guise of intuition. And there has
been a truly astonishing outcome of studying infinite sets. Their study led to the
discovery of fundamental, logical limits on what computers can possibly do. For
example, in Section 8.2, we’ll use reasoning developed for infinite sets to prove
that it’s impossible to have a perfect type-checker for a programming language.
So in this chapter, we ask you to bite the bullet and start learning to cope with
infinity.
“mcs” — 2017/6/5 — 19:42 — page 258 — #266
258
Chapter 8
Infinite Sets
8.1
Infinite Cardinality
In the late nineteenth century, the mathematician Georg Cantor was studying the
convergence of Fourier series and found some series that he wanted to say converged “most of the time,” even though there were an infinite number of points
where they didn’t converge. As a result, Cantor needed a way to compare the size
of infinite sets. To get a grip on this, he got the idea of extending the Mapping Rule
Theorem 4.5.4 to infinite sets: he regarded two infinite sets as having the “same
size” when there was a bijection between them. Likewise, an infinite set A should
be considered “as big as” a set B when A surj B. So we could consider A to be
“strictly smaller” than B, which we abbreviate as A strict B, when A is not “as big
as” B:
Definition 8.1.1.
A strict B
iff
NOT .A
surj B/.
On finite sets, this strict relation really does mean “strictly smaller.” This follows
immediately from the Mapping Rule Theorem 4.5.4.
Corollary 8.1.2. For finite sets A; B,
A strict B
iff
jAj < jBj:
Proof.
A strict B
iff
NOT .A
surj B/
iff
NOT .jAj
jBj/
(Def 8.1.1)
(Theorem 4.5.4.(4.5))
iff jAj < jBj:
Cantor got diverted from his study of Fourier series by his effort to develop a
theory of infinite sizes based on these ideas. His theory ultimately had profound
consequences for the foundations of mathematics and computer science. But Cantor made a lot of enemies in his own time because of his work: the general mathematical community doubted the relevance of what they called “Cantor’s paradise”
of unheard-of infinite sizes.
A nice technical feature of Cantor’s idea is that it avoids the need for a definition
of what the “size” of an infinite set might be—all it does is compare “sizes.”
Warning: We haven’t, and won’t, define what the “size” of an infinite set is.
The definition of infinite “sizes” requires the definition of some infinite sets called
“mcs” — 2017/6/5 — 19:42 — page 259 — #267
8.1. Infinite Cardinality
259
ordinals with special well-ordering properties. The theory of ordinals requires getting deeper into technical set theory than we need to go, and we can get by just
fine without defining infinite sizes. All we need are the “as big as” and “same size”
relations, surj and bij, between sets.
But there’s something else to watch out for: we’ve referred to surj as an “as big
as” relation and bij as a “same size” relation on sets. Of course, most of the “as big
as” and “same size” properties of surj and bij on finite sets do carry over to infinite
sets, but some important ones don’t—as we’re about to show. So you have to be
careful: don’t assume that surj has any particular “as big as” property on infinite
sets until it’s been proven.
Let’s begin with some familiar properties of the “as big as” and “same size”
relations on finite sets that do carry over exactly to infinite sets:
Lemma 8.1.3. For any sets A; B; C ,
1. A surj B iff B inj A.
2. If A surj B and B surj C , then A surj C .
3. If A bij B and B bij C , then A bij C .
4. A bij B iff B bij A.
Part 1. follows from the fact that R has the Œ 1 out; 1 in surjective function
property iff R 1 has the Œ 1 out; 1 in total, injective property. Part 2. follows
from the fact that compositions of surjections are surjections. Parts 3. and 4. follow from the first two parts because R is a bijection iff R and R 1 are surjective
functions. We’ll leave verification of these facts to Problem 4.21.
Another familiar property of finite sets carries over to infinite sets, but this time
some real ingenuity is needed to prove it:
Theorem 8.1.4. [Schröder-Bernstein] For any sets A; B, if A surj B and B surj A,
then A bij B.
That is, the Schröder-Bernstein Theorem says that if A is at least as big as B
and conversely, B is at least as big as A, then A is the same size as B. Phrased
this way, you might be tempted to take this theorem for granted, but that would be
a mistake. For infinite sets A and B, the Schröder-Bernstein Theorem is actually
pretty technical. Just because there is a surjective function f W A ! B—which
need not be a bijection—and a surjective function g W B ! A—which also need
not be a bijection—it’s not at all clear that there must be a bijection e W A ! B. The
idea is to construct e from parts of both f and g. We’ll leave the actual construction
to Problem 8.12.
“mcs” — 2017/6/5 — 19:42 — page 260 — #268
260
Chapter 8
Infinite Sets
Another familiar set property is that for any two sets, either the first is at least
as big as the second, or vice-versa. For finite sets this follows trivially from the
Mapping Rule. It’s actually still true for infinite sets, but assuming it is obvious
would be mistaken again.
Theorem 8.1.5. For all sets A; B,
A surj B
OR
B surj A:
Theorem 8.1.5 lets us prove that another basic property of finite sets carries over
to infinite ones:
Lemma 8.1.6.
A strict B AND B strict C
(8.1)
implies
A strict C
for all sets A; B; C .
Proof. (of Lemma 8.1.6)
Suppose 8.1 holds, and assume for the sake of contradiction that NOT.A strict
C /, which means that A surj C . Now since B strict C , Theorem 8.1.5 lets us
conclude that C surj B. So we have
A surj C AND C surj B;
and Lemma 8.1.3.2 lets us conclude that A surj B, contradicting the fact that
A strict B.
We’re omitting a proof of Theorem 8.1.5 because proving it involves technical
set theory—typically the theory of ordinals again—that we’re not going to get into.
But since proving Lemma 8.1.6 is the only use we’ll make of Theorem 8.1.5, we
hope you won’t feel cheated not to see a proof.
8.1.1
Infinity is different
A basic property of finite sets that does not carry over to infinite sets is that adding
something new makes a set bigger. That is, if A is a finite set and b … A, then
jA [ fbgj D jAj C 1, and so A and A [ fbg are not the same size. But if A is
infinite, then these two sets are the same size!
Lemma 8.1.7. Let A be a set and b … A. Then A is infinite iff A bij A [ fbg.
“mcs” — 2017/6/5 — 19:42 — page 261 — #269
8.1. Infinite Cardinality
261
Proof. Since A is not the same size as A [ fbg when A is finite, we only have to
show that A [ fbg is the same size as A when A is infinite.
That is, we have to find a bijection between A [ fbg and A when A is infinite.
Here’s how: since A is infinite, it certainly has at least one element; call it a0 . But
since A is infinite, it has at least two elements, and one of them must not equal to
a0 ; call this new element a1 . But since A is infinite, it has at least three elements,
one of which must not equal both a0 and a1 ; call this new element a2 . Continuing
in this way, we conclude that there is an infinite sequence a0 ; a1 ; a2 ; : : : ; an ; : : : of
different elements of A. Now it’s easy to define a bijection e W A [ fbg ! A:
e.b/ WWD a0 ;
e.an / WWD anC1
e.a/ WWD a
for n 2 N;
for a 2 A
fb; a0 ; a1 ; : : :g:
8.1.2
Countable Sets
A set C is countable iff its elements can be listed in order, that is, the elements in
C are precisely the elements in the sequence
c0 ; c1 ; : : : ; cn ; : : : :
Assuming no repeats in the list, saying that C can be listed in this way is formally
the same as saying that the function, f W N ! C defined by the rule that f .i /WWDci ,
is a bijection.
Definition 8.1.8. A set C is countably infinite iff N bij C . A set is countable iff it
is finite or countably infinite. A set is uncountable iff it is not countable.
We can also make an infinite list using just a finite set of elements if we allow
repeats. For example, we can list the elements in the three-element set f2; 4; 6g as
2; 4; 6; 6; 6; : : : :
This simple observation leads to an alternative characterization of countable sets
that does not make separate cases of finite and infinite sets. Namely, a set C is
countable iff there is a list
c0 ; c1 ; : : : ; cn ; : : :
of the elements of C , possibly with repeats.
Lemma 8.1.9. A set C is countable iff N surj C . In fact, a nonempty set C is
countable iff there is a total surjective function g W N ! C .
“mcs” — 2017/6/5 — 19:42 — page 262 — #270
262
Chapter 8
Infinite Sets
The proof is left to Problem 8.13.
The most fundamental countably infinite set is the set N itself. But the set Z of
all integers is also countably infinite, because the integers can be listed in the order:
0; 1; 1; 2; 2; 3; 3; : : : :
(8.2)
In this case, there is a simple formula for the nth element of the list (8.2). That is,
the bijection f W N ! Z such that f .n/ is the nth element of the list can be defined
as:
(
n=2
if n is even;
f .n/ WWD
.n C 1/=2 if n is odd:
There is also a simple way to list all pairs of nonnegative integers, which shows
that .N N/ is also countably infinite (Problem 8.23). From this, it’s a small
step to reach the conclusion that the set Q0 of nonnegative rational numbers is
countable. This may be a surprise—after all, the rationals densely fill up the space
between integers, and for any two, there’s another in between. So it might seem as
though you couldn’t write out all the rationals in a list, but Problem 8.11 illustrates
how to do it. More generally, it is easy to show that countable sets are closed under
unions and products (Problems 8.22 and 8.23) which implies the countability of a
bunch of familiar sets:
Corollary 8.1.10. The following sets are countably infinite:
ZC ; Z; N N; QC ; Z Z; Q:
A small modification of the proof of Lemma 8.1.7 shows that countably infinite
sets are the “smallest” infinite sets. Namely,
Lemma 8.1.11. If A is an infinite set, and B is countable, then A surj B.
We leave the proof to Problem 8.10.
Also, since adding one new element to an infinite set doesn’t change its size,
you can add any finite number of elements without changing the size by simply
adding one element after another. Something even stronger is true: you can add a
countably infinite number of new elements to an infinite set and still wind up with
just a set of the same size (Problem 8.15).
By the way, it’s a common mistake to think that, because you can add any finite
number of elements to an infinite set and have a bijection with the original set, that
you can also throw in infinitely many new elements. In general it isn’t true that just
because it’s OK to do something any finite number of times, it’s also OK to do it an
infinite number of times. For example, starting from 3, you can increment by 1 any
finite number of times, and the result will be some integer greater than or equal to
3. But if you increment an infinite number of times, you don’t get an integer at all.
“mcs” — 2017/6/5 — 19:42 — page 263 — #271
8.1. Infinite Cardinality
8.1.3
263
Power sets are strictly bigger
Cantor’s astonishing discovery was that not all infinite sets are the same size. In
particular, he proved that for any set A the power set pow.A/ is “strictly bigger”
than A. That is,
Theorem 8.1.12. [Cantor] For any set A,
A strict pow.A/:
Proof. To show that A is strictly smaller than pow.A/, we have to show that if g is
a function from A to pow.A/, then g is not a surjection. Since any partial function
with nonempty codomain can be extended to a total function with the same range
(reader: ask yourself how), we can safely assume that g is total.
To show that g is not a surjection, we’ll simply find a subset Ag A that is not
in the range of g. The idea is, for any element a 2 A, to look at the set g.a/ A
and ask whether or not a happens to be in g.a/. First, define
Ag WWD fa 2 A j a … g.a/g:
Ag is a well-defined subset of A, which means it is a member of pow.A/. But Ag
can’t be in the range of g, because it differs at a from each set g.a/ in the range of
g.
To spell this out more, suppose to the contrary that Ag was in the range of g, that
is,
Ag D g.a0 /
for some a0 2 A. Now by definition of Ag ,
a 2 g.a0 / iff a 2 Ag
iff a … g.a/
for all a 2 A. Now letting a D a0 yields the contradiction
a0 2 g.a0 / iff a0 … g.a0 /:
So g is not a surjection, because there is an element in the power set of A, specifically the set Ag , that is not in the range of g.
Cantor’s Theorem immediately implies:
Corollary 8.1.13. pow.N/ is uncountable.
Proof. By Lemma 8.1.9, U is uncountable iff N strict U .
“mcs” — 2017/6/5 — 19:42 — page 264 — #272
264
Chapter 8
Infinite Sets
The bijection between subsets of an n-element set and the length n bit-strings
f0; 1gn used to prove Theorem 4.5.5, carries over to a bijection between subsets of
a countably infinite set and the infinite bit-strings, f0; 1g! . That is,
pow.N/ bij f0; 1g! :
This immediately implies
Corollary 8.1.14. f0; 1g! is uncountable.
More Countable and Uncountable Sets
Once we have a few sets we know are countable or uncountable, we can get lots
more examples using Lemma 8.1.3. In particular, we can appeal to the following
immediate corollary of the Lemma:
Corollary 8.1.15.
(a) If U is an uncountable set and A surj U , then A is uncountable.
(b) If C is a countable set and C surj A, then A is countable.
For example, now that we know that the set f0; 1g! of infinite bit strings is uncountable, it’s a small step to conclude that
Corollary 8.1.16. The set R of real numbers is uncountable.
To prove this, think about the infinite decimal expansion of a real number:
p
2 D 1:4142 : : : ;
5 D 5:000 : : : ;
1=10 D 0:1000 : : : ;
1=3 D 0:333 : : : ;
1=9 D 0:111 : : : ;
1
D 4:010101 : : : :
4
99
Let’s map any real number r to the infinite bit string b.r/ equal to the sequence
of bits in the decimal expansion of r, starting at the decimal point. If the decimal
expansion of r happens to contain a digit other than 0 or 1, leave b.r/ undefined.
“mcs” — 2017/6/5 — 19:42 — page 265 — #273
8.1. Infinite Cardinality
265
For example,
b.5/ D 000 : : : ;
b.1=10/ D 1000 : : : ;
b.1=9/ D 111 : : : ;
1
b.4 / D 010101 : : :
99
p
b. 2/; b.1=3/ are undefined:
Now b is a function from real numbers to infinite bit strings.1 It is not a total
function, but it clearly is a surjection. This shows that
R surj f0; 1g! ;
and the uncountability of the reals now follows by Corollary 8.1.15.(a).
For another example, let’s prove
Corollary 8.1.17. The set .ZC / of all finite sequences of positive integers is countable.
To prove this, think about the prime factorization of a nonnegative integer:
20 D 22 30 51 70 110 130 ;
6615 D 20 33 51 72 110 130 :
Let’s map any nonnegative integer n to the finite sequence e.n/ of nonzero exponents in its prime factorization. For example,
e.20/ D .2; 1/;
e.6615/ D .3; 1; 2/;
e.5
13
9
817
11 47
10344 / D .13; 9; 817; 44/;
e.1/ D ;
(the empty string)
e.0/ is undefined:
1 Some
rational numbers can be expanded in two ways—as an infinite sequence ending in all 0’s
or as an infinite sequence ending in all 9’s. For example,
5 D 5:000 D 4:999 : : : ;
1
D 0:1000 D 0:0999 : : : :
10
In such cases, define b.r/ to be the sequence that ends with all 0’s.
“mcs” — 2017/6/5 — 19:42 — page 266 — #274
266
Chapter 8
Infinite Sets
Now e is a function from N to .ZC / . It is defined on all positive integers, and it
clearly is a surjection. This shows that
N surj .ZC / ;
and the countability of the finite strings of positive integers now follows by Corollary 8.1.15.(b).
Larger Infinities
There are lots of different sizes of infinite sets. For example, starting with the
infinite set N of nonnegative integers, we can build the infinite sequence of sets
N strict pow.N/ strict pow.pow.N// strict pow.pow.pow.N/// strict : : : :
By Cantor’s Theorem 8.1.12, each of these sets is strictly bigger than all the preceding ones. But that’s not all: the union of all the sets in the sequence is strictly
bigger than each set in the sequence (see Problem 8.14). In this way you can keep
going indefinitely, building “bigger” infinities all the way.
8.1.4
Diagonal Argument
Theorem 8.1.12 and similar proofs are collectively known as “diagonal arguments”
because of a more intuitive version of the proof described in terms of on an infinite
square array. Namely, suppose there was a bijection between N and f0; 1g! . If such
a relation existed, we would be able to display it as a list of the infinite bit strings
in some countable order or another. Once we’d found a viable way to organize
this list, any given string in f0; 1g! would appear in a finite number of steps, just
as any integer you can name will show up a finite number of steps from 0. This
hypothetical list would look something like the one below, extending to infinity
both vertically and horizontally:
A0
A1
A2
A3
A4
A5
::
:
D
D
D
D
D
D
1
0
1
0
0
1
::
:
0
1
1
1
0
0
::
:
0
1
1
0
1
0
::
:
0
1
1
0
0
1
::
:
1
0
1
1
0
1
::
:
1
1
1
0
0
1
::
:
::
:
“mcs” — 2017/6/5 — 19:42 — page 267 — #275
8.2. The Halting Problem
267
But now we can exhibit a sequence that’s missing from our allegedly complete list
of all the sequences. Look at the diagonal in our sample list:
A0
A1
A2
A3
A4
A5
::
:
D
D
D
D
D
D
1
0
1
0
0
1
::
:
0
1
1
1
0
0
::
:
0
1
1
0
1
0
::
:
0
1
1
0
0
1
::
:
1
0
1
1
0
1
::
:
1
1
1
0
0
1
::
:
::
:
Here is why the diagonal argument has its name: we can form a sequence D consisting of the bits on the diagonal.
D D 1 1 1 0 0 1 ;
Then, we can form another sequence by switching the 1’s and 0’s along the diagonal. Call this sequence C :
C D 0 0 0 1 1 0 :
Now if the nth term of An is 1 then the nth term of C is 0, and vice versa, which
guarantees that C differs from An . In other words, C has at least one bit different
from every sequence on our list. So C is an element of f0; 1g! that does not appear
in our list—our list can’t be complete!
This diagonal sequence C corresponds to the set fa 2 A j a … g.a/g in the
proof of Theorem 8.1.12. Both are defined in terms of a countable subset of the
uncountable infinity in a way that excludes them from that subset, thereby proving
that no countable subset can be as big as the uncountable set.
8.2
The Halting Problem
Although towers of larger and larger infinite sets are at best a romantic concern for
a computer scientist, the reasoning that leads to these conclusions plays a critical
role in the theory of computation. Diagonal arguments are used to show that lots of
problems can’t be solved by computation, and there is no getting around it.
This story begins with a reminder that having procedures operate on programs
is a basic part of computer science technology. For example, compilation refers to
taking any given program text written in some “high level” programming language
“mcs” — 2017/6/5 — 19:42 — page 268 — #276
268
Chapter 8
Infinite Sets
like Java, C++, Python, . . . , and then generating a program of low-level instructions that does the same thing but is targeted to run well on available hardware.
Similarly, interpreters or virtual machines are procedures that take a program text
designed to be run on one kind of computer and simulate it on another kind of computer. Routine features of compilers involve “type-checking” programs to ensure
that certain kinds of run-time errors won’t happen, and “optimizing” the generated
programs so they run faster or use less memory.
The fundamental thing that just can’t be done by computation is a perfect job of
type-checking, optimizing, or any kind of analysis of the overall run time behavior
of programs. In this section, we’ll illustrate this with a basic example known as
the Halting Problem. Pick your favorite programming language—Python, Java,
C++,. . . —and assume that “program” refers to a program written in your language.
Once a program is started, if its initial computation stops for some reason—such
as producing a final value, waiting for an input, suffering an error interrupt, or simply freezing—the program is said to halt. So a program that does not halt would run
forever using up cycles and energy (unless it was interrupted by an external operating system command). The Halting Problem is the general problem of determining,
given an arbitrary program, whether or not the program halts.
There is a simple way to determine when an arbitrary program does halt, at least
in theory: just run it until it stops. Well not quite. Just running the program in the
usual way would not detect when it freezes without warning. What really needs to
be done is to simulate the program using an interpreter or a virtual machine that
will recognize any kind of halting, including a freeze. But interpreters and virtual
machines capable of simulating any program are familiar technology. So there is a
general way to detect when a program halts.
The hard part is determining when a program does not halt. At any point in simulating it, you generally won’t know whether to continue the simulation because
the program will halt later, or to abort the simulation because it wouldn’t stop otherwise.
So is there some way besides simulation to detect when a program does not
halt? Could there be some program analysis tool that could inspect any program
and correctly report when the program does not halt? The answer is “No.” Using
a standard diagonal argument, we will prove that it is impossible to have such a
non-halting analysis tool. Any method for detecting non-halting is bound to go
wrong. It will falsely report that some halting program does not halt, or it will fail
to report anything about some program. That is, there will be a program on which
the analyzer runs forever without halting.
To set up the diagonal argument, we will focus on string procedures. A string
procedure is a procedure that takes a single argument that is supposed to be a string
“mcs” — 2017/6/5 — 19:42 — page 269 — #277
8.2. The Halting Problem
269
over the 256 character ASCII alphabet. As a simple example, you might think about
how to write a string procedure that halts precisely when it is applied to a double
letter string in ASCII , namely, a string in which every character occurs twice in a
row. For example, aaCC33, and zz++ccBB are double letter strings, but aa;bb,
b33, and AAAAA are not.
When the computation of a procedure applied to a string halts, the procedure will
be said to recognize the string. In this context, a set of strings a commonly called a
(formal) language. We let lang.P / to be the language recognized by procedure P :
lang.P / WWDfs 2 ASCII j P applied to s haltsg:
A language is called recognizable when it equals lang.P / for some string procedure P . For example, we’ve just agreed that the set of double letter strings is
recognizable.
There is no harm in assuming that every program can be written as a string in
ASCII —that’s typically how we would enter them into a computer in the first
place. When a string s 2 ASCII is actually the ASCII description of some string
procedure, we’ll refer to that string procedure as Ps . You can think of Ps as the
result of compiling s into something executable.2 It’s technically helpful to treat
every string in ASCII as a program for a string procedure. So when a string
s 2 ASCII doesn’t parse as a proper string procedure, we’ll define Ps to be some
default string procedure—say one that never halts on anything it is applied to.
Focusing just on string procedures, the general Halting Problem is to decide,
given strings s and t , whether or not the procedure Ps recognizes t . Following the
usual diagonal approach, we define the language No-halt:
Definition 8.2.1.
No-halt WWD fs j Ps applied to s does not haltg D fs … lang.Ps /g:
(8.3)
So if Ps is the recognizer for some language, then No-halt differs from that language on the string s. This shows that No-halt cannot have a recognizer:
Theorem 8.2.2. No-halt is not recognizable.
Let’s spell out the reasoning behind this theorem more fully. By definition, we
have
s 2 No-halt IFF s … lang.Ps /;
(8.4)
string s 2 ASCII and the procedure Ps have to be distinguished to avoid a type error:
you can’t apply a string to string. For example, let s be the string that you wrote as your program
to recognize the double letter strings. Applying s to a string argument, say aabbccdd, should
throw a type exception; what you need to do is compile s to the procedure Ps and then apply Ps to
aabbccdd.
2 The
“mcs” — 2017/6/5 — 19:42 — page 270 — #278
270
Chapter 8
Infinite Sets
for all strings s 2 ASCII .
Now suppose to the contrary that No-halt was recognizable. This means there is
some procedure Ps0 that recognizes No-halt, that is,
No-halt D lang.Ps0 / :
Combined with (8.4), we have
s 2 lang.Ps0 /
iff s … lang.Ps /
(8.5)
for all s 2 ASCII . Now letting s D s0 in (8.5) yields the immediate contradiction
s0 2 lang.Ps0 /
iff s0 … lang.Ps0 / :
So that does it: the reasoning above applied to whatever favorite programming
language you picked. It is logically impossible for a Java program to be a recognizer
for non-halting Java programs, or for a Python program to be a recognizer for nonhalting Python programs, and so forth for other favorite programming languages.
Now you might wonder if there was a loophole around this logical limitation
by having a recognizer the non-halting programs in one language that was written
in another language. In other words, could there be a C++ procedure that recognizes all the non-halting Java programs? After all, C++ does allow more intimate
manipulation of computer memory than Java does. But there is no loophole here.
If you’ve learned about programming language implementation, you will realize
that it’s possible to write a simulator in Java for C++ programs. This means that if
there were a C++ procedure that recognized non-halting Java programs, then a Java
procedure could also do it by simulating the C++ program, and that’s impossible.
This reasoning finally leads to a transcendent insight. No procedure can be written
in any programming language that recognizes No-halt for your favorite language.
Recognizing No-halt is simply beyond the capacity of computation.
But it’s not just No-halt. If there was a perfect recognizer for any property that
depends on the complete run time behavior programs,3 it could be altered slightly
to become a recognizer for No-halt. We’ll take this claim for granted now, giving
its full justification in an assigned problem.
For example, most compilers do “static” type-checking at compile time to ensure
that programs won’t make run-time type errors. A program that type-checks is
guaranteed not to cause a run-time type-error. But given that it’s impossible to
recognize when the complete run time behavior of a program does not lead to a
3 The weasel words “complete run time behavior” creep in here to rule out some run time properties that are easy to recognize because they depend only on part of the run time behavior. For
example, the set of programs that halt after executing at most 100 instructions is recognizable.
“mcs” — 2017/6/5 — 19:42 — page 271 — #279
8.3. The Logic of Sets
271
type-error, it follows that the type-checker must be rejecting programs that really
wouldn’t cause a type-error. The conclusion is that no type-checker is perfect—you
can always do better!
It’s a different story if we think about the practical possibility of writing program analyzers. The fact that it’s logically impossible to analyze utterly arbitrary
programs does not mean that you can’t do a very good job analyzing interesting programs that come up in practice. In fact, these “interesting” programs are commonly
intended to be analyzable in order to confirm that they do what they’re supposed to
do.
In the end, it’s not clear how much of a hurdle this theoretical limitation implies
in practice. But the theory does provide some perspective on claims about general
analysis methods for programs. The theory tells us that people who make such
claims either
are exaggerating the power (if any) of their methods, perhaps to make a sale
or get a grant, or
are trying to keep things simple by not going into technical limitations they’re
aware of, or
perhaps most commonly, are so excited about some useful practical successes
of their methods that they haven’t bothered to think about the limitations
which must be there.
So from now on, if you hear people making claims about having general program
analysis/verification/optimization methods, you’ll know they can’t be telling the
whole story.
8.3
The Logic of Sets
8.3.1
Russell’s Paradox
Reasoning naively about sets turns out to be risky. In fact, one of the earliest attempts to come up with precise axioms for sets in the late nineteenth century by
the logician Gotlob Frege, was shot down by a three line argument known as Russell’s Paradox4 which reasons in nearly the same way as the proof of Cantor’s
4 Bertrand
Russell was a mathematician/logician at Cambridge University at the turn of the Twentieth Century. He reported that when he felt too old to do mathematics, he began to study and write
about philosophy, and when he was no longer smart enough to do philosophy, he began writing about
politics. He was jailed as a conscientious objector during World War I. For his extensive philosophical
and political writing, he won a Nobel Prize for Literature.
“mcs” — 2017/6/5 — 19:42 — page 272 — #280
272
Chapter 8
Infinite Sets
Theorem 8.1.12. This was an astonishing blow to efforts to provide an axiomatic
foundation for mathematics:
Russell’s Paradox
Let S be a variable ranging over all sets, and define
W WWD fS j S 62 S g:
So by definition,
S 2 W iff S 62 S;
for every set S . In particular, we can let S be W , and obtain the
contradictory result that
W 2 W iff W 62 W:
The simplest reasoning about sets crashes mathematics! Russell and his colleague Whitehead spent years trying to develop a set theory that was not contradictory, but would still do the job of serving as a solid logical foundation for all of
mathematics.
Actually, a way out of the paradox was clear to Russell and others at the time:
it’s unjustified to assume that W is a set. The step in the proof where we let S be
W has no justification, because S ranges over sets, and W might not be a set. In
fact, the paradox implies that W had better not be a set!
But denying that W is a set means we must reject the very natural axiom that
every mathematically well-defined collection of sets is actually a set. The problem faced by Frege, Russell and their fellow logicians was how to specify which
well-defined collections are sets. Russell and his Cambridge University colleague
Whitehead immediately went to work on this problem. They spent a dozen years
developing a huge new axiom system in an even huger monograph called Principia Mathematica, but for all intents and purposes, their approach failed. It was
so cumbersome no one ever used it, and it was subsumed by a much simpler, and
now widely accepted, axiomatization of set theory by the logicians Zermelo and
Fraenkel.
“mcs” — 2017/6/5 — 19:42 — page 273 — #281
8.3. The Logic of Sets
8.3.2
273
The ZFC Axioms for Sets
A formula of set theory5 is a predicate formula that only talks about membership
in sets. That is, a first-order formula of set theory is built using logical connectives
and quantifiers starting solely from expressions of the form “x 2 y.” The domain
of discourse is the collection of sets, and “x 2 y” is interpreted to mean that x and
y are variables that range over sets, and x is one of the elements in y.
Formulas of set theory are not even allowed to have the equality symbol “D,” but
sets are equal iff they have the same elements, so there is an easy way to express
equality of sets purely in terms of membership:
.x D y/ WWD 8z: .z 2 x IFF z 2 y/:
(8.6)
Similarly, the subset symbol “” is not allowed in formulas of set theory, but we
can also express subset purely in terms of membership:
.x y/ WWD 8z: .z 2 x IMPLIES z 2 y/:
(8.7)
So formulas using symbols “D; ,” in addition to “2” can be understood as
abbreviations for formulas only using “2.” We won’t worry about this distinction
between formulas and abbreviations for formulas—we’ll now just call them all
“formulas of set theory.” For example,
x D y IFF Œx y AND y x
is a formula of set theory that explains a basic connection between set equality and
set containment.
It’s generally agreed that essentially all of mathematics can be derived from a
few formulas of set theory, called the Axioms of Zermelo-Fraenkel Set Theory with
Choice (ZFC), using a few simple logical deduction rules.
We’re not going to be studying the axioms of ZFC in this text, but we thought
you might like to see them—and while you’re at it, get some more practice reading
and writing quantified formulas:
Extensionality. Two sets are equal iff they are members of the same sets:
x D y IFF .8z: x 2 z IFF y 2 z/:
Pairing. For any two sets x and y, there is a set fx; yg with x and y as its only
elements:
8x; y9u8z: Œz 2 u IFF .z D x OR z D y/
5 Technically
this is called a pure first-order formula of set theory
“mcs” — 2017/6/5 — 19:42 — page 274 — #282
274
Chapter 8
Infinite Sets
Union. The union u of a collection z of sets is also a set:
8z9u8x: .x 2 u/ IFF .9y: x 2 y AND y 2 z/
Infinity. There is an infinite set. Specifically, there is a nonempty set x such that
for any set y 2 x, the set fyg is also a member of x.
Subset. Given any set x and any definable property of sets, there is a set y containing precisely those elements in x that have the property.
8x9y8z: z 2 y IFF Œz 2 x AND .z/
where .z/ is a formula of set theory.6
Power Set. All the subsets of a set form another set:
8x9p8u: u x IFF u 2 p:
Replacement. Suppose a formula of set theory defines the graph of a function
on a set s, that is,
8x 2 s 8y; z: Œ.x; y/ AND .x; z/ IMPLIES y D z:
Then the image of s under that function is also a set t. Namely,
9t8y: y 2 t IFF Œ9x 2 s: .x; y/:
Foundation. The aim is to forbid any infinite sequence of sets of the form
2 xn 2 2 x1 2 x0
in which each set is a member of the next one. This can be captured by saying
every nonempty set has a “member-minimal” element. Namely, define
member-minimal.m; x/ WWD Œm 2 x AND 8y 2 x: y … m:
Then the Foundation Axiom7 is
8x: x ¤ ; IMPLIES 9m: member-minimal.m; x/:
Choice. Let s be a set of nonempty, disjoint sets. Then there is a set c consisting of
exactly one element from each set in s. The formula is given in Problem 8.35.
6 This
7 This
axiom is more commonly called the Comprehension Axiom.
axiom is also called the Regularity Axiom.
“mcs” — 2017/6/5 — 19:42 — page 275 — #283
8.4. Does All This Really Work?
8.3.3
275
Avoiding Russell’s Paradox
These modern ZFC axioms for set theory are much simpler than the system Russell
and Whitehead first came up with to avoid paradox. In fact, the ZFC axioms are
as simple and intuitive as Frege’s original axioms, with one technical addition: the
Foundation axiom. Foundation captures the intuitive idea that sets must be built up
from “simpler” sets in certain standard ways. And in particular, Foundation avoids
the “paradox” of Russell’s set collection W WWD fS j S … S g because it implies that
W is not a set. Namely Foundation implies that S 62 S for every sets S , so W is
the collection of all sets. Now if W was a set, we would have W 2 W , violating
Foundation.
8.4
Does All This Really Work?
So this is where mainstream mathematics stands today: there is a handful of ZFC
axioms from which virtually everything else in mathematics can be derived logically. This sounds like a rosy situation, but there are several dark clouds, suggesting
that the essence of truth in mathematics is not completely resolved.
The ZFC axioms weren’t etched in stone by God. Instead, they were mostly
made up by Zermelo, who may have been a brilliant logician, but was also
a fallible human being—probably some days he forgot his house keys. So
maybe Zermelo, just like Frege, didn’t get his axioms right and will be
shot down by some successor to Russell who will use his axioms to prove
a proposition P and its negation P . Then math as we understand it would be
broken—this may sound crazy, but it has happened before.
In fact, while there is broad agreement that the ZFC axioms are capable of
proving all of standard mathematics, the axioms have some further consequences that sound paradoxical. For example, the Banach-Tarski Theorem
says that, as a consequence of the axiom of choice, a solid ball can be divided
into six pieces and then the pieces can be rigidly rearranged to give two solid
balls of the same size as the original!
Some basic questions about the nature of sets remain unresolved. For example, Cantor raised the question whether there is a set whose size is strictly
between the smallest infinite set N (see Problem 8.10) and the strictly larger
set pow.N/? Cantor guessed not:
“mcs” — 2017/6/5 — 19:42 — page 276 — #284
276
Chapter 8
Infinite Sets
Cantor’s Contiuum Hypothesis: There is no set A such that
N strict A strict pow.N/:
The Continuum Hypothesis remains an open problem a century later. Its
difficulty arises from one of the deepest results in modern Set Theory—
discovered in part by Gödel in the 1930’s and Paul Cohen in the 1960’s—
namely, the ZFC axioms are not sufficient to settle the Continuum Hypothesis: there are two collections of sets, each obeying the laws of ZFC, and in
one collection the Continuum Hypothesis is true, and in the other it is false.
Until a mathematician with a deep understanding of sets can extend ZFC with
persuasive new axioms, the Continuum Hypothesis will remain undecided.
But even if we use more or different axioms about sets, there are some unavoidable problems. In the 1930’s, Gödel proved that, assuming that an axiom system like ZFC is consistent—meaning you can’t prove both P and P
for any proposition, P —then the very proposition that the system is consistent (which is not too hard to express as a logical formula) cannot be proved
in the system. In other words, no consistent system is strong enough to verify
itself.
8.4.1
Large Infinities in Computer Science
If the romance of different-size infinities and continuum hypotheses doesn’t appeal
to you, not knowing about them is not going to limit you as a computer scientist.
These abstract issues about infinite sets rarely come up in mainstream mathematics, and they don’t come up at all in computer science, where the focus is generally
on “countable,” and often just finite, sets. In practice, only logicians and set theorists have to worry about collections that are “too big” to be sets. That’s part of
the reason that the 19th century mathematical community made jokes about “Cantor’s paradise” of obscure infinities. But the challenge of reasoning correctly about
this far-out stuff led directly to the profound discoveries about the logical limits of
computation described in Section 8.2, and that really is something every computer
scientist should understand.
Problems for Section 8.1
Practice Problems
Problem 8.1.
Show that the set f0; 1g of finite binary strings is countable.
“mcs” — 2017/6/5 — 19:42 — page 277 — #285
8.4. Does All This Really Work?
277
Problem 8.2.
Describe an example of two uncountable sets A and B such that there is no bijection between A and B.
Problem 8.3.
Indicate which of the following assertions (there may be more than one) are equivalent to
A strict N:
jAj is undefined.
A is countably infinite.
A is uncountable.
A is finite.
N surj A.
8n 2 N, jAj n.
8n 2 N, jAj n.
9n 2 N: jAj n.
9n 2 N: jAj < n.
Problem 8.4.
Prove that if there is a total injective (Œ 1 out; 1 in) relation from S to N, then
S is countable.
Problem 8.5.
Prove that if S is an infinite set, then pow.S / is uncountable.
Problem 8.6.
Let A to be some infinite set. We know from Lemma 8.1.7 that
A bij .A [ fb0 g/
(Aub)
“mcs” — 2017/6/5 — 19:42 — page 278 — #286
278
Chapter 8
Infinite Sets
for any element b0 . An easy induction implies that
A bij .A [ fb0 ; b1 ; : : : ; bn g/
(Aubi )
for any finite set fb0 ; b1 ; : : : ; bn g.
Students sometimes assume that this implies that
A bij .A [ B/;
(8.8)
but it doesn’t. For example, (AuB) is not true if A is N and B is the real numbers
R.8
A collection C of sets is called a chain when, given any two sets in C, one is a
subset of the other. A predicate P .C / is said to be finitely continuousS
if, whenever
F is a chain of finite sets, and P .F / is true for every F 2 F, then P . F/ is true.
Claiming that Aubi implies (AuB) amounts to claiming that the predicate PA .C /
is finitely continuous, where
PA .C / WWD A bij .A [ C /:
But it isn’t, as the example with A D N and C D R demonstrates.
Briefly explain which of the following predicates P .C / is finitely continuous
and which not.
1. C is finite.
2. C is uncountable.
3. C D ;.
4. There is a minimum element b 2 C \ N.
5. There is a minimum element b 2 C \ Z.
6. =2 2 C .
7. 9 > 0 8a; b 2 C \ R: ja
bj > .
8. C [ N is finite.
9. C N.
10. C N.
8 It happens that(AuB) is true if B
and takes some proving.
is countable, (Problem 8.15), but this is not completely obvious
“mcs” — 2017/6/5 — 19:42 — page 279 — #287
8.4. Does All This Really Work?
279
11. C is countable.
Problem 8.7.
Prove that the set of all finite subsets of positive integers is countable. Hint: Classify the subsets by the sum of their elements.
Problem 8.8.
A collection C of sets is called a chain when, given any two
S sets in C, one is a subset
of the other. Prove that if F is chain of finite sets, then F is countable. (Notice
that without the chain condition, every set is the union of its finite subsets.)
Class Problems
Problem 8.9.
Show that the set N of finite sequences of nonnegative integers is countable.
Problem 8.10. (a) Several students felt the proof of Lemma 8.1.7 was worrisome,
if not circular. What do you think?
(b) Use the proof of Lemma 8.1.7 to show that if A is an infinite set, then A surj N,
that is, every infinite set is “at least as big as” the set of nonnegative integers.
Problem 8.11.
The rational numbers fill the space between integers, so a first thought is that there
must be more of them than the integers, but it’s not true. In this problem you’ll
show that there are the same number of positive rationals as positive integers. That
is, the positive rationals are countable.
(a) Define a bijection between the set ZC of positive integers, and the set .ZC ZC / of all pairs of positive integers:
.1; 1/; .1; 2/; .1; 3/; .1; 4/; .1; 5/; : : :
.2; 1/; .2; 2/; .2; 3/; .2; 4/; .2; 5/; : : :
.3; 1/; .3; 2/; .3; 3/; .3; 4/; .3; 5/; : : :
.4; 1/; .4; 2/; .4; 3/; .4; 4/; .4; 5/; : : :
.5; 1/; .5; 2/; .5; 3/; .5; 4/; .5; 5/; : : :
::
:
“mcs” — 2017/6/5 — 19:42 — page 280 — #288
280
Chapter 8
Infinite Sets
(b) Conclude that the set QC of all positive rational numbers is countable.
Problem 8.12.
This problem provides a proof of the [Schröder-Bernstein] Theorem:
If A inj B and B inj A, then A bij B.
(8.9)
Since A inj B and B inj A, there are are total injective functions f W A ! B
and g W B ! A.
Assume for simplicity that A and B have no elements in common. Let’s picture
the elements of A arranged in a column, and likewise B arranged in a second column to the right, with left-to-right arrows connecting a to f .a/ for each a 2 A
and likewise right-to-left arrows for g. Since f and g are total functions, there is
exactly one arrow out of each element. Also, since f and g are injections, there is
at most one arrow into any element.
So starting at any element, there is a unique and unending path of arrows going
forwards (it might repeat). There is also a unique path of arrows going backwards,
which might be unending, or might end at an element that has no arrow into it.
These paths are completely separate: if two ran into each other, there would be two
arrows into the element where they ran together.
This divides all the elements into separate paths of four kinds:
(i) paths that are infinite in both directions,
(ii) paths that are infinite going forwards starting from some element of A.
(iii) paths that are infinite going forwards starting from some element of B.
(iv) paths that are unending but finite.
(a) What do the paths of the last type (iv) look like?
(b) Show that for each type of path, either
(i) the f -arrows define a bijection between the A and B elements on the path, or
(ii) the g-arrows define a bijection between B and A elements on the path, or
(iii) both sets of arrows define bijections.
For which kinds of paths do both sets of arrows define bijections?
(c) Explain how to piece these bijections together to form a bijection between A
and B.
“mcs” — 2017/6/5 — 19:42 — page 281 — #289
8.4. Does All This Really Work?
281
(d) Justify the assumption that A and B are disjoint.
Problem 8.13. (a) Prove that if a nonempty set C is countable, then there is a total
surjective function f W N ! C .
(b) Conversely, suppose that N surj D, that is, there is a not necessarily total
surjective function f W ND. Prove that D is countable.
Problem 8.14.
There are lots of different sizes of infinite sets. For example, starting with the
infinite set N of nonnegative integers, we can build the infinite sequence of sets
N strict pow.N/ strict pow.pow.N// strict pow.pow.pow.N/// strict : : : :
where each set is “strictly smaller” than the next one by Theorem 8.1.12. Let
pown .N/ be the nth set in the sequence, and
U WWD
1
[
pown .N/:
nD0
(a) Prove that
U surj pown .N/;
(8.10)
for all n > 0.
(b) Prove that
pown .N/ strict U
for all n 2 N.
Now of course, we could take U; pow.U /; pow.pow.U //; : : : and keep on in this
way building still bigger infinities indefinitely.
Homework Problems
Problem 8.15.
Prove that if A is an infinite set and B is a countable, then
A bij .A [ B/:
Hint: Since A is infinite, we can find an infinite sequence a0 ; a1 ; a2 ; : : : of distinct elements of A as in the proof of Lemma 8.1.7.
“mcs” — 2017/6/5 — 19:42 — page 282 — #290
282
Chapter 8
Infinite Sets
Problem 8.16.
In this problem you will prove a fact that may surprise you—or make you even
more convinced that set theory is nonsense: the half-open unit interval is actually
the “same size” as the nonnegative quadrant of the real plane!9 Namely, there is a
bijection from .0; 1 to Œ0; 1/ Œ0; 1/.
(a) Describe a bijection from .0; 1 to Œ0; 1/.
Hint: 1=x almost works.
(b) An infinite sequence of the decimal digits f0; 1; : : : ; 9g will be called long if
it does not end with all 0’s. An equivalent way to say this is that a long sequence
is one that has infinitely many occurrences of nonzero digits. Let L be the set of
all such long sequences. Describe a bijection from L to the half-open real interval
.0; 1.
Hint: Put a decimal point at the beginning of the sequence.
(c) Describe a surjective function from L to L2 that involves alternating digits
from two long sequences. Hint: The surjection need not be total.
(d) Prove the following lemma and use it to conclude that there is a bijection from
L2 to .0; 12 .
Lemma 8.4.1. Let A and B be nonempty sets. If there is a bijection from A to B,
then there is also a bijection from A A to B B.
(e) Conclude from the previous parts that there is a surjection from .0; 1 to .0; 12 .
Then appeal to the Schröder-Bernstein Theorem to show that there is actually a
bijection from .0; 1 to .0; 12 .
(f) Complete the proof that there is a bijection from .0; 1 to Œ0; 1/2 .
Problem 8.17.
You don’t really have to go down the diagonal in a “diagonal” argument.
Let’s review the historic application of a diagonal argument to one-way infinite
sequences
he0 ; e1 ; e2 ; : : : ; ek ; : : : i :
The angle brackets appear above as a reminder that the sequence is not a set: its
elements appear in order, and the same element may appear multiple times.10
9 The
half-open unit interval .0; 1 is fr 2 R j 0 < r 1g. Similarly, Œ0; 1/ WWD fr 2 R j r 0g.
If such
10 The right angle-bracket is not really visible, since the sequence does not have a right end.
“mcs” — 2017/6/5 — 19:42 — page 283 — #291
8.4. Does All This Really Work?
283
The general setup for a diagonal argument is that we have some sequence S
whose elements are themselves one-way infinite sequences. We picture the sequence S D hs0 ; s1 ; s2 ; : : : i running vertically downward, and each sequence
sk 2 S running horizontally to the right:
˝
˛
sk D sk;0 ; sk;1 ; sk;2 ; : : : :
So we have a 2-D matrix that is infinite down and to the right:
0
1
2
s0 s0;0 s0;1 s0;2
s1 s1;0 s1;1 s1;2
s2 s2;0 s2;1 s2;2
::
:
sk
:::
:::
:::
:::
k :::
s0;k : : :
s1;k : : :
s2;k : : :
::
:
:::
sk:k : : :
The diagonal argument explains how to find a “new” sequence, that is, a sequence that is not in S. Namely, create the new sequence by going down the diagonal of the matrix and, for each element encountered, add a differing element to the
sequence being created. In other words, the diagonal sequence is
˝
˛
DS WWD s0;0 ; s1;1 ; s2;2 ; : : : ; sk;k ; : : :
where x indicates some element that is not equal to x. Now DS is a sequence that
is not in S because it differs from every sequence in S , namely, it differs from the
kth sequence in S at position k.
For definiteness, let’s say
(
1 if x ¤ 1;
x WWD
2 otherwise:
With this contrivance, we have gotten the diagonal sequence DS to be a sequence
of 1’s and 2’s that is not in S .
But as we said at the beginning, you don’t have to go down the diagonal. You
could, for example, follow a line with a slope of 1=4 to get a new sequence
˝
˛
TS WWD s0;0 ; t1 ; t2 ; t3 ; s1;4 ; t5 ; t6 ; t7 ; s2;8 ; : : : ; t4k 1 ; sk;4k ; t4kC1 ; : : :
one-way infinite sequences seem worrisome, you can replace them with total functions on N. So the
sequence above simpy becomes the function e on N where e.n/ WWD en .
“mcs” — 2017/6/5 — 19:42 — page 284 — #292
284
Chapter 8
Infinite Sets
where the ti ’s can be anything at all. Any such Ts will be a new sequence because
it will differ from every row of the matrix, but this time it differs from the kth row
at position 4k.
(a) By letting all the ti ’s be 2’s, we get a new sequence TS whose elements (in
the limit) are at least three-quarters equal to 2. Explain how to find an uncountable
number of such sequences.
Hint: Use slope 1=8 and fill six out of eight places with 2’s. The abbreviation
2.n/ WWD 2; 2; : : : ; 2
„ ƒ‚ …
length n
for a length-n sequence of 2’s may be helpful, in particular 2.6/ .
(b) Let’s say a sequence has a negligible fraction of non-2 elements if, in the limit,
it has a fraction of at most non-2 elements for every > 0. Describe how to
define a sequence not in S that has a negligible fraction of non-2 elements.
(c) Describe how to find a sequence that differs infinitely many times from every
sequence in S .
Hint: Divide N into an infinite number of non-overlapping infinite pieces.
Problem 8.18.
An infinite binary string b WWD b0 b1 b2 : : : bn : : : is called OK when the 1’s in b
occur only at perfect square positions. That is, b is OK when
(
0
if i … f0; 1; 4; 9; : : : ; n2 ; .n C 1/2 : : :g;
bi D
0 or 1 otherwise:
(a) Prove that the set of OK strings is uncountable.
(b) Prove that a set with an uncountable subset must itself be uncountable.
(c) Let Sparse be the set of infinite binary strings whose fraction of 1’s approaches
zero. Conclude that Sparse is uncountable.
Problem 8.19.
Cantor’s Powerset Theorem 8.1.12 implies that for any set A, no total function
“mcs” — 2017/6/5 — 19:42 — page 285 — #293
8.4. Does All This Really Work?
285
g W A ! pow.A/ is a surjection. In particular, the theorem is proved by showing
that the “diagonal” set
Ag WWD fa 2 A j a … g.a/g
is an element of pow.A/ that is not in the range of g, because for every a 2 A it
differs at the element a from the set g.a/ in the range of g.
But there’s no need to stick to the diagonal. If p W A ! A is some total function,
we can differ from g.a/ at p.a/ instead of “on the diagonal” at a. That is, define
Ag;p WWD fp.a/ j a 2 A AND p.a/ … g.a/g:
So Ag;p is an element of pow.A/, and
False Claim. Ag;p is not in the range of g,
because for every a 2 A, the set Ag;p differs at the element p.a/ from the set
g.a/ in the range of g.
(a) Show that the claim is false by letting A be f0; 1g, and g.n/ D fng and p.n/ D
0 for all n 2 N.
(b) Identify the mistake in the argument above and show that it is fixed by requiring that p be an injection.
Problem 8.20.
Suppose S D fS0 ; S1 ; : : :g is a countable set each of whose elements Sn is an
infinite set of nonnegative integers. Using a diagonal argument, we can find a
“new” infinite set U of nonnegative integers that is not in S.
In this problem we describe how to find an infinite set U of nonnegative integers
that not only is not in S, but does not even have a subset that is in S.
Rather than describing U directly, it’s a little easier if we describe its complement
C . Then U will be C WWD N C .
Define a function f W N ! N as follows:
f .n/ WWD minfk 2 Sn j k n2 g;
and define
C WWD range.f /:
(a) Prove that no subset of U is in S.
(b) Show that the limiting density of U is one. That is,
jU \ Œ0::kj
D 1:
k
k!1
lim
“mcs” — 2017/6/5 — 19:42 — page 286 — #294
286
Chapter 8
Infinite Sets
Exam Problems
Problem 8.21. (a) For each of the following sets, indicate whether it is finite,
countably infinite, or uncountable.
(i) The set of even integers greater than 10100 .
(ii) The set of “pure” complex numbers of the form ri for nonzero real numbers
r.
(iii) The powerset of the integer interval Œ10::1010 .
(iv) The complex numbers c such that c is the root of a quadratic with integer
coefficients, that is,
9m; n; p 2 Z; m ¤ 0: mc 2 C nc C p D 0:
Let U be an uncountable set, C be a countably infinite subset of U, and D be
a countably infinite set.
(v) U [ D.
(vi) U \ C
(vii) U D
(b) Give an example of sets A and B such that
R strict A strict B:
Problem 8.22.
Prove that if A0 ; A1 ; : : : ; An ; : : : is an infinite sequence of countable sets, then so
is
1
[
An
nD0
Problem 8.23.
Let A and B be countably infinite sets:
A D fa0 ; a1 ; a2 ; a3 ; : : :g
B D fb0 ; b1 ; b2 ; b3 ; : : :g
Show that their product A B is also a countable set by showing how to list the
elements of AB. You need only show enough of the initial terms in your sequence
to make the pattern clear—a half dozen or so terms usually suffice.
“mcs” — 2017/6/5 — 19:42 — page 287 — #295
8.4. Does All This Really Work?
287
Problem 8.24.
Let f0; 1g be the set of finite binary sequences, f0; 1g! be the set of infinite binary sequences, and F be the set of sequences in f0; 1g! that contain only a finite
number of occurrences of 1’s.
(a) Describe a simple surjective function from f0; 1g to F .
(b) The set F WWD f0; 1g! F consists of all the infinite binary sequences with
infinitely many 1’s. Use the previous problem part to prove that F is uncountable.
Hint: We know that f0; 1g is countable and f0; 1g! is not.
Problem 8.25.
Let f0; 1g! be the set of infinite binary strings, and let B f0; 1g! be the set of
infinite binary strings containing infinitely many occurrences of 1’s. Prove that B
is uncountable. (We have already shown that f0; 1g! is uncountable.)
Hint: Start by showing that f0; 1g! inj B.
Problem 8.26.
A real number is called quadratic when it is a root of a degree two polynomial with
integer coefficients. Explain why there are only countably many quadratic reals.
Problem 8.27.
Describe which of the following sets have bijections between them:
Z (integers);
C (complex numbers);
pow.Z/ (all subsets of integers);
pow.pow.;//;
f0; 1g! (infinite binary sequences)
pow.fT; Fg/;
R (real numbers);
Q (rational numbers);
pow.;/;
f0; 1g (finite binary sequences);
fT; Fg (truth values)
pow.f0; 1g! /
Problem 8.28.
Prove that the set ZC of all finite sequences of positive integers is countable.
Hint: If s 2 ZC , let sum(s) be the sum of the successive integers in s.
“mcs” — 2017/6/5 — 19:42 — page 288 — #296
288
Chapter 8
Infinite Sets
Problem 8.29.
Let f1; 2; 3g! be the set of infinite sequences containing only the numbers 1, 2, and
3. For example, some sequences of this kind are:
.1; 1; 1; 1:::/;
.2; 2; 2; 2:::/;
.3; 2; 1; 3:::/:
Prove that f1; 2; 3g! is uncountable.
Hint: One approach is to define a surjective function from f1; 2; 3g! to the power
set pow.N/.
Problems for Section 8.2
Class Problems
Problem 8.30.
Let N! be the set of infinite sequences of nonnegative integers. For example, some
sequences of this kind are:
.0; 1; 2; 3; 4; : : : /;
.2; 3; 5; 7; 11; : : : /;
.3; 1; 4; 5; 9; : : : /:
Prove that this set of sequences is uncountable.
Homework Problems
Problem 8.31.
For any sets A and B, let ŒA ! B be the set of total functions from A to B. Prove
that if A is not empty and B has more than one element, then NOT.A surj ŒA !
B/.
Hint: Suppose that is a function from A to ŒA ! B mapping each element
a 2 A to a function a W A ! B. Pick any two elements of B; call them 0 and 1.
Then define
(
0 if a .a/ D 1;
diag.a/ WWD
1 otherwise:
“mcs” — 2017/6/5 — 19:42 — page 289 — #297
8.4. Does All This Really Work?
289
Problem 8.32.
String procedures are one-argument procedures that apply to strings over the ASCII
alphabet. If application of procedure P to string s results in a computation that
eventually halts, we say that P recognizes s. We define lang.P / to be the set of
strings or language recognized by P :
lang.P / WWDfs 2 ASCII j P recognizes sg:
A language is unrecognizable when it is not equal to lang.P / for any procedure P .
A string procedure declaration is a text s 2 ASCII that conforms to the grammatical rules for programs. The declaration defines a procedure Ps , which we can
think of as the result of compiling s into an executable object. If s 2 ASCII is not
a grammatically well-formed procedure declaration, we arbitrarily define Ps to be
the string procedure that fails to halt when applied to any string. Now every string
defines a string procedure, and every string procedure is Ps for some s 2 ASCII .
An easy diagonal argument in Section 8.2 showed that
No-halt WWD fs j Ps applied to s does not haltg D fs j s … lang.Ps /g
is not recognizable.
It may seem pretty weird to apply a procedure to its own declaration. Are there
any less weird examples of unrecognizable set? The answer is “many more.” In
this problem, we’ll show three more:
No-halt- WWD fs j Ps applied to does not haltg D fs j … lang.Ps /g;
Finite-halt WWD fs j lang.Ps / is finiteg;
Always-halt WWD fs j lang.Ps / D ASCII g:
Let’s begin by showing how we could use a recognizer for No-halt- to define a
recognizer for No-halt. That is, we will “reduce” the weird problem of recognizing
No-halt to the more understandable problem of recognizing No-halt-. Since there
is no recognizer for No-halt, it follows that there can’t be one for No-halt- either.
Here’s how this reduction would work: suppose we want to recognize when a
given string s is in No-halt. Revise s to be the declaration of a slightly modified
procedure Ps 0 which behaves as follows:
Ps 0 applied to argument t 2 ASCII , ignores t, and simulates Ps applied to s.
So, if Ps applied to s halts, then Ps 0 halts on every string it is applied to, and if Ps
applied to s does not halt, then Ps 0 does not halt on any string it is applied to. That
“mcs” — 2017/6/5 — 19:42 — page 290 — #298
290
Chapter 8
Infinite Sets
is,
s 2 No-halt IMPLIES lang.Ps 0 / D ;
IMPLIES
… lang.Ps 0 /
IMPLIES
s 0 2 No-halt-;
s … No-halt IMPLIES lang.Ps 0 / D ASCII
IMPLIES
2 lang.Ps 0 /
IMPLIES
s 0 … No-halt-:
In short,
s 2 No-halt IFF s 0 2 No-halt-:
So to recognize when s 2 No-halt all you need to do is recognize when s 0 2
No-halt-. As already noted above (but we know that remark got by several students,
so we’re repeating the explanation), this means that if No-halt- was recognizable,
then No-halt would be as well. Since we know that No-halt is unrecognizable, then
No-halt- must also be unrecognizable, as claimed.
(a) Conclude that Finite-halt is unrecognizable.
Hint: Same s 0 .
Next, let’s see how a reduction of No-halt to Always-halt would work. Suppose we
want to recognize when a given string s is in No-halt. Revise s to be the declaration
of a slightly modified procedure Ps 00 which behaves as follows:
When Ps 00 is applied to argument t 2 ASCII , it simulates Ps applied
to s for up to jt j “steps” (executions of individual machine instructions). If Ps applied to s has not halted in jtj steps, then the application
of Ps 00 to t halts. If Ps applied to s has halted within jtj steps, then the
application of Ps 00 to t runs forever.
(b) Conclude that Always-halt is unrecognizable.
Hint: Explain why
s 2 No-halt IFF s 00 2 Always-halt:
(c) Explain why Finite-halt is unrecognizable.
Hint: Same s 00 .
Note that it’s easy to recognize when Ps does halt on s: just simulate the application of Ps to s until it halts. This shows that No-halt is recognizable. We’ve just
concluded that Finite-halt is nastier: neither it nor its complement is recognizable.
“mcs” — 2017/6/5 — 19:42 — page 291 — #299
8.4. Does All This Really Work?
291
Exam Problems
Problem 8.33.
Let f1; 2; 3g! be the set of infinite sequences containing only the numbers 1, 2, and
3. For example, some sequences of this kind are:
.1; 1; 1; 1:::/;
.2; 2; 2; 2:::/;
.3; 2; 1; 3:::/:
Prove that f1; 2; 3g! is uncountable.
Hint: One approach is to define a surjective function from f1; 2; 3g! to the power
set pow.N/.
Problems for Section 8.3
Class Problems
Problem 8.34.
Forming a pair .a; b/ of items a and b is a mathematical operation that we can
safely take for granted. But when we’re trying to show how all of mathematics can
be reduced to set theory, we need a way to represent the pair .a; b/ as a set.
(a) Explain why representing .a; b/ by fa; bg won’t work.
(b) Explain why representing .a; b/ by fa; fbgg won’t work either. Hint: What
pair does ff1g; f2gg represent?
(c) Define
pair.a; b/ WWD fa; fa; bgg:
Explain why representing .a; b/ as pair.a; b/ uniquely determines a and b. Hint:
Sets can’t be indirect members of themselves: a 2 a never holds for any set a, and
neither can a 2 b 2 a hold for any b.
Problem 8.35.
The axiom of choice says that if s is a set whose members are nonempty sets that
are pairwise disjoint—that is, no two sets in s have an element in common—then
there is a set c consisting of exactly one element from each set in s.
“mcs” — 2017/6/5 — 19:42 — page 292 — #300
292
Chapter 8
Infinite Sets
In formal logic, we could describe s with the formula,
pairwise-disjoint.s/WWD8x 2 s: x ¤ ; AND
8x; y 2 s: x ¤ y IMPLIES x \ y D ;:
Similarly we could describe c with the formula
choice-set.c; s/ WWD
8x 2 s: 9Šz: z 2 c \ x:
Here “9Š z:” is fairly standard notation for “there exists a unique z.”
Now we can give the formal definition:
Definition (Axiom of Choice).
8s: pairwise-disjoint.s/ IMPLIES 9c: choice-set.c; s/:
The only issue here is that set theory is technically supposed to be expressed in
terms of pure formulas in the language of sets, which means formula that uses only
the membership relation 2 propositional connectives, the two quantifies 8 and 9,
and variables ranging over all sets. Verify that the axiom of choice can be expressed
as a pure formula, by explaining how to replace all impure subformulas above with
equivalent pure formulas.
For example, the formula x D y could be replaced with the pure formula 8z: z 2
x IFF z 2 y.
Problem 8.36.
Let R W A ! A be a binary relation on a set A. If a1 R a0 , we’ll say that a1 is “Rsmaller” than a0 . R is called well founded when there is no infinite “R-decreasing”
sequence:
R an R R a1 R a0 ;
(8.11)
of elements ai 2 A.
For example, if A D N and R is the <-relation, then R is well founded because
if you keep counting down with nonnegative integers, you eventually get stuck at
zero:
0 < < n 1 < n:
But you can keep counting up forever, so the >-relation is not well founded:
> n > > 1 > 0:
Also, the -relation on N is not well founded because a constant sequence of, say,
2’s, gets -smaller forever:
2 2 2:
“mcs” — 2017/6/5 — 19:42 — page 293 — #301
8.4. Does All This Really Work?
293
(a) If B is a subset of A, an element b 2 B is defined to be R-minimal in B iff
there is no R-smaller element in B. Prove that R W A ! A is well founded iff every
nonempty subset of A has an R-minimal element.
A logic formula of set theory has only predicates of the form “x 2 y” for variables x; y ranging over sets, along with quantifiers and propositional operations.
For example,
isempty.x/ WWD 8w: NOT.w 2 x/
is a formula of set theory that means that “x is empty.”
(b) Write a formula member-minimal.u; v/ of set theory that means that u is 2minimal in v.
(c) The Foundation axiom of set theory says that 2 is a well founded relation
on sets. Express the Foundation axiom as a formula of set theory. You may use
“member-minimal” and “isempty” in your formula as abbreviations for the formulas defined above.
(d) Explain why the Foundation axiom implies that no set is a member of itself.
Homework Problems
Problem 8.37.
In writing formulas, it is OK to use abbreviations introduced earlier (so it is now
legal to use “D” because we just defined it).
(a) Explain how to write a formula, Subsetn .x; y1 ; y2 ; : : : ; yn /, of set theory 11
that means x fy1 ; y2 ; : : : ; yn g.
(b) Now use the formula Subsetn to write a formula, Atmostn .x/, of set theory
that means that x has at most n elements.
(c) Explain how to write a formula Exactlyn of set theory that means that x has
exactly n elements. Your formula should only be about twice the length of the
formula Atmostn .
(d) The direct way to write a formula Dn .y1 ; : : : ; yn / of set theory that means
that y1 ; : : : ; yn are distinct elements is to write an AND of subformulas “yi ¤ yj ”
for 1 i < j n. Since there are n.n 1/=2 such subformulas, this approach
leads to a formula Dn whose length grows proportional to n2 . Describe how to
write such a formula Dn .y1 ; : : : ; yn / whose length only grows proportional to n.
Hint: Use Subsetn and Exactlyn .
11 See
Section 8.3.2.
“mcs” — 2017/6/5 — 19:42 — page 294 — #302
294
Chapter 8
Infinite Sets
Exam Problems
Problem 8.38. (a) Explain how to write a formula Members.p; a; b/ of set theory12 that means p D fa; bg.
Hint: Say that everything in p is either a or b. It’s OK to use subformulas of the
form “x D y,” since we can regard “x D y” as an abbreviation for a genuine set
theory formula.
A pair .a; b/ is simply a sequence of length two whose first item is a and whose
second is b. Sequences are a basic mathematical data type we take for granted, but
when we’re trying to show how all of mathematics can be reduced to set theory, we
need a way to represent the ordered pair .a; b/ as a set. One way that will work13
is to represent .a; b/ as
pair.a; b/ WWD fa; fa; bgg:
(b) Explain how to write a formula Pair.p; a; b/, of set theory 14 that means p D
pair.a; b/.
Hint: Now it’s OK to use subformulas of the form “Members.p; a; b/.”
(c) Explain how to write a formula Second.p; b/, of set theory that means p is a
pair whose second item is b.
Problems for Section 8.4
Homework Problems
Problem 8.39.
In this problem, structural induction and the Foundation Axiom of set theory provide simple proofs about some utterly infinite objects.
Definition. The class of ‘recursive set-like” objects, Recs, is defined recursively as
follows:
Base case: The empty set ; is a Recs.
Constructor step: If S is a nonempty set of Recs’s, then S is a Recs.
12 See
Section 8.3.2.
similar ways that don’t work are described in problem 8.34.
14 See Section 8.3.2.
13 Some
“mcs” — 2017/6/5 — 19:42 — page 295 — #303
8.4. Does All This Really Work?
295
(a) Prove that Recs satisfies the Foundation Axiom: there is no infinite sequence
of Recs, ro ; r1 ; : : : ; rn 1 ; rn ; : : : such that
: : : rn 2 rn
1
2 : : : r1 2 r0 :
(8.12)
Hint: Structural induction.
(b) Prove that every pure set is a Recs.15
Hint: Use the Foundation axiom.
(c) Every Recs R defines a special kind of two-person game of perfect information called a uniform game. The initial “board position” of the game is R itself.
A player’s move consists of choosing any member R. The two players alternate
moves, with the player whose turn it is to move called the Next player. The Next
player’s move determines a game in which the other player, called the Previous
player, moves first.
The game is called “uniform” because the two players have the same objective: to
leave the other player stuck with no move to make. That is, whoever moves to the
empty set is a winner, because then the next player has no move.
Prove that in every uniform game, either the Previous player or the Next player has
a winning strategy.
Problem 8.40.
For any set x, define next.x/ to be the set consisting of all the elements of x, along
with x itself:
next.x/ WWD x [ fxg:
So by definition,
x 2 next.x/ and x next.x/:
(8.13)
Now we give a recursive definition of a collection Ord of sets called ordinals that
provide a way to count infinite sets. Namely,
Definition.
; 2 Ord;
if 2 Ord; then next./ 2 Ord;
[
if S Ord; then
2 Ord:
2S
15 A
“pure” set is empty or is a set whose elements are all pure sets.
“mcs” — 2017/6/5 — 19:42 — page 296 — #304
296
Chapter 8
Infinite Sets
There is a method for proving things about ordinals that follows directly from
the way they are defined. Namely, let P .x/ be some property of sets. The Ordinal
Induction Rule says that to prove that P ./ is true for all ordinals , you need only
show two things
If P holds for all the members of next.x/, then it holds for next.x/, and
if P holds for all members of some set S , then it holds for their union.
That is:
Rule. Ordinal Induction
8x: .8y 2 next.x/: P .y// IMPLIES S
P .next.x//;
8S: .8x 2 S: P .x// IMPLIES P . x2S x/
8 2 Ord: P ./
The intuitive justification for the Ordinal Induction Rule is similar to the justification for strong induction. We will accept the soundness of the Ordinal Induction
Rule as a basic axiom.
(a) A set x is closed under membership if every element of x is also a subset of
x, that is
8y 2 x: y x:
Prove that every ordinal is closed under membership.
(b) A sequence
2 nC1 2 n 2 2 1 2 0
(8.14)
of ordinals i is called a member-decreasing sequence starting at 0 . Use Ordinal
Induction to prove that no ordinal starts an infinite member-decreasing sequence.16
16 Do not assume the Foundation Axiom of ZFC (Section 8.3.2) which says that there isn’t any set
that starts an infinite member-decreasing sequence. Even in versions of set theory in which the Foundation Axiom does not hold, there cannot be any infinite member-decreasing sequence of ordinals.
“mcs” — 2017/6/5 — 19:42 — page 297 — #305
II
Structures
“mcs” — 2017/6/5 — 19:42 — page 298 — #306
“mcs” — 2017/6/5 — 19:42 — page 299 — #307
Introduction
The properties of the set of integers are the subject of Number Theory. This part
of the text starts with a chapter on this topic because the integers are a very familiar mathematical structure that have lots of easy-to-state and interesting-to-prove
properties. This makes Number Theory a good place to start serious practice with
the methods of proof outlined in Part 1. Moreover, Number Theory has turned out
to have multiple applications in computer science. For example, most modern data
encryption methods are based on Number theory.
We study numbers as a “structure” that has multiple parts of different kinds. One
part is, of course, the set of all the integers. A second part is the collection of basic
integer operations: addition, multiplication, exponentiation,. . . . Other parts are the
important subsets of integers—like the prime numbers—out of which all integers
can be built using multiplication.
Structured objects more generally are fundamental in computer science. Whether
you are writing code, solving an optimization problem, or designing a network, you
will be dealing with structures.
Graphs, also known as networks, are a fundamental structure in computer science. Graphs can model associations between pairs of objects; for example, two
exams that cannot be given at the same time, two people that like each other, or two
subroutines that can be run independently. In Chapter 10, we study directed graphs
which model one-way relationships such as being bigger than, loving (sadly, it’s
often not mutual), and being a prerequisite for. A highlight is the special case of
acyclic digraphs (DAGs) that correspond to a class of relations called partial orders. Partial orders arise frequently in the study of scheduling and concurrency.
Digraphs as models for data communication and routing problems are the topic of
Chapter 11.
In Chapter 12 we focus on simple graphs that represent mutual or symmetric re-
“mcs” — 2017/6/5 — 19:42 — page 300 — #308
300
Part II
Structures
lationships, such as being in conflict, being compatible, being independent, being
capable of running in parallel. Planar Graphs—simple graphs that can be drawn in
the plane—are examined in Chapter 13, the final chapter of Part II. The impossibility of placing 50 geocentric satellites in orbit so that they uniformly blanket the
globe will be one of the conclusions reached in this chapter.
“mcs” — 2017/6/5 — 19:42 — page 301 — #309
9
Number Theory
Number theory is the study of the integers. Why anyone would want to study the
integers may not be obvious. First of all, what’s to know? There’s 0, there’s 1, 2,
3, and so on, and, oh yeah, -1, -2, . . . . Which one don’t you understand? What
practical value is there in it?
The mathematician G. H. Hardy delighted at its impracticality. He wrote:
[Number theorists] may be justified in rejoicing that there is one science, at any rate, and that their own, whose very remoteness from ordinary human activities should keep it gentle and clean.
Hardy was especially concerned that number theory not be used in warfare; he
was a pacifist. You may applaud his sentiments, but he got it wrong: number theory
underlies modern cryptography, which is what makes secure online communication
possible. Secure communication is of course crucial in war—leaving poor Hardy
spinning in his grave. It’s also central to online commerce. Every time you buy a
book from Amazon, use a certificate to access a web page, or use a PayPal account,
you are relying on number theoretic algorithms.
Number theory also provides an excellent environment for us to practice and
apply the proof techniques that we developed in previous chapters. We’ll work out
properties of greatest common divisors (gcd’s) and use them to prove that integers
factor uniquely into primes. Then we’ll introduce modular arithmetic and work out
enough of its properties to explain the RSA public key crypto-system.
Since we’ll be focusing on properties of the integers, we’ll adopt the default
convention in this chapter that variables range over the set Z of integers.
9.1
Divisibility
The nature of number theory emerges as soon as we consider the divides relation.
Definition 9.1.1. a divides b (notation a j b) iff there is an integer k such that
ak D b:
The divides relation comes up so frequently that multiple synonyms for it are
used all the time. The following phrases all say the same thing:
“mcs” — 2017/6/5 — 19:42 — page 302 — #310
302
Chapter 9
Number Theory
a j b,
a divides b,
a is a divisor of b,
a is a factor of b,
b is divisible by a,
b is a multiple of a.
Some immediate consequences of Definition 9.1.1 are that for all n
n j 0;
n j n; and
˙ 1 j n:
Also,
0 j n IMPLIES n D 0:
Dividing seems simple enough, but let’s play with this definition. The Pythagoreans, an ancient sect of mathematical mystics, said that a number is perfect if it
equals the sum of its positive integral divisors, excluding itself. For example,
6 D 1 C 2 C 3 and 28 D 1 C 2 C 4 C 7 C 14 are perfect numbers. On the
other hand, 10 is not perfect because 1 C 2 C 5 D 8, and 12 is not perfect because
1 C 2 C 3 C 4 C 6 D 16. Euclid characterized all the even perfect numbers around
300 BC (Problem 9.2). But is there an odd perfect number? More than two thousand years later, we still don’t know! All numbers up to about 10300 have been
ruled out, but no one has proved that there isn’t an odd perfect number waiting just
over the horizon.
So a half-page into number theory, we’ve strayed past the outer limits of human
knowledge. This is pretty typical; number theory is full of questions that are easy to
pose, but incredibly difficult to answer. We’ll mention a few more such questions
in later sections.1
9.1.1
Facts about Divisibility
The following lemma collects some basic facts about divisibility.
Lemma 9.1.2.
1. If a j b and b j c, then a j c.
1 Don’t
Panic—we’re going to stick to some relatively benign parts of number theory. These
super-hard unsolved problems rarely get put on problem sets.
“mcs” — 2017/6/5 — 19:42 — page 303 — #311
9.1. Divisibility
303
2. If a j b and a j c, then a j sb C t c for all s and t .
3. For all c ¤ 0, a j b if and only if ca j cb.
Proof. These facts all follow directly from Definition 9.1.1. To illustrate this, we’ll
prove just part 2:
Given that a j b, there is some k1 2 Z such that ak1 D b. Likewise, ak2 D c,
so
sb C t c D s.k1 a/ C t .k2 a/ D .sk1 C t k2 /a:
Therefore sb C t c D k3 a where k3 WWD .sk1 C t k2 /, which means that
a j sb C t c:
A number of the form sb C t c is called an integer linear combination of b and c,
or, since in this chapter we’re only talking about integers, just a linear combination.
So Lemma 9.1.2.2 can be rephrased as
If a divides b and c, then a divides every linear combination of b and c.
We’ll be making good use of linear combinations, so let’s get the general definition
on record:
Definition 9.1.3. An integer n is a linear combination of numbers b0 ; : : : ; bk iff
n D s0 b0 C s1 b1 C C sk bk
for some integers s0 ; : : : ; sk .
9.1.2
When Divisibility Goes Bad
As you learned in elementary school, if one number does not evenly divide another,
you get a “quotient” and a “remainder” left over. More precisely:
Theorem 9.1.4. [Division Theorem]2 Let n and d be integers such that d ¤ 0.
Then there exists a unique pair of integers q and r, such that
n D q d C r AND 0 r < jd j :
2 This
(9.1)
theorem is often called the “Division Algorithm,” but we prefer to call it a theorem since it
does not actually describe a division procedure for computing the quotient and remainder.
“mcs” — 2017/6/5 — 19:42 — page 304 — #312
304
Chapter 9
Number Theory
The number q is called the quotient and the number r is called the remainder of
n divided by d . We use the notation qcnt.n; d / for the quotient and rem.n; d / for
the remainder.
The absolute value notation jd j used above is probably familiar from introductory calculus, but for the record, let’s define it.
Definition 9.1.5. For any real number r, the absolute value jrj of r is:3
(
r
if r 0;
jrj WWD
r if r < 0:
So by definition, the remainder rem.n; d / is nonnegative regardless of the sign
of n and d . For example, rem. 11; 7/ D 3, since 11 D . 2/ 7 C 3.
“Remainder” operations built into many programming languages can be a source
of confusion. For example, the expression “32 % 5” will be familiar to programmers in Java, C, and C++; it evaluates to rem.32; 5/ D 2 in all three languages.
On the other hand, these and other languages are inconsistent in how they treat remainders like “32 % -5” or “-32 % 5” that involve negative numbers. So don’t be
distracted by your familiar programming language’s behavior on remainders, and
stick to the mathematical convention that remainders are nonnegative.
The remainder on division by d by definition is a number in the (integer) interval
from 0 to jd j 1. Such integer intervals come up so often that it is useful to have a
simple notation for them. For k n 2 Z,
9.1.3
.k::n/ WWD
fi j k < i < ng;
.k::n WWD
.k; n/ [ fng;
Œk::n/ WWD
fkg [ .k; n/;
Œk::n WWD
fkg [ .k; n/ [ fng D fi j k i ng:
Die Hard
Die Hard 3 is just a B-grade action movie, but we think it has an inner message:
everyone should learn at least a little number theory. In Section 6.2.3, we formalized a state machine for the Die Hard jug-filling problem using 3 and 5 gallon jugs,
p
absolute value of r could be defined as r 2 , which works because of the convention that
square root notation always refers to the nonnegative square root (see Problem 1.3). Absolute value
generalizes to complex numbers where it is called the norm. For a; b 2 R,
p
ja C bi j WWD a2 C b 2 :
3 The
“mcs” — 2017/6/5 — 19:42 — page 305 — #313
9.1. Divisibility
305
and also with 3 and 9 gallon jugs, and came to different conclusions about bomb
explosions. What’s going on in general? For example, how about getting 4 gallons
from 12- and 18-gallon jugs, getting 32 gallons with 899- and 1147-gallon jugs, or
getting 3 gallons into a jug using just 21- and 26-gallon jugs?
It would be nice if we could solve all these silly water jug questions at once. This
is where number theory comes in handy.
A Water Jug Invariant
Suppose that we have water jugs with capacities a and b with b a. Let’s carry
out some sample operations of the state machine and see what happens, assuming
the b-jug is big enough:
.0; 0/ ! .a; 0/
fill first jug
! .0; a/
pour first into second
! .a; a/
fill first jug
! .2a
b; b/
pour first into second (assuming 2a b)
! .2a
b; 0/
empty second jug
! .0; 2a
b/
pour first into second
! .a; 2a
b/
fill first
! .3a
2b; b/
pour first into second (assuming 3a 2b)
What leaps out is that at every step, the amount of water in each jug is a linear
combination of a and b. This is easy to prove by induction on the number of
transitions:
Lemma 9.1.6 (Water Jugs). In the Die Hard state machine of Section 6.2.3 with
jugs of sizes a and b, the amount of water in each jug is always a linear combination
of a and b.
Proof. The induction hypothesis P .n/ is the proposition that after n transitions, the
amount of water in each jug is a linear combination of a and b.
Base case (n D 0): P .0/ is true, because both jugs are initially empty, and 0 a C
0 b D 0.
Inductive step: Suppose the machine is in state .x; y/ after n steps, that is, the little
jug contains x gallons and the big one contains y gallons. There are two cases:
If we fill a jug from the fountain or empty a jug into the fountain, then that jug
is empty or full. The amount in the other jug remains a linear combination
of a and b. So P .n C 1/ holds.
“mcs” — 2017/6/5 — 19:42 — page 306 — #314
306
Chapter 9
Number Theory
Otherwise, we pour water from one jug to another until one is empty or the
other is full. By our assumption, the amount x and y in each jug is a linear
combination of a and b before we begin pouring. After pouring, one jug is
either empty (contains 0 gallons) or full (contains a or b gallons). Thus, the
other jug contains either x C y, x C y a or x C y b gallons, all of which
are linear combinations of a and b since x and y are. So P .n C 1/ holds in
this case as well.
Since P .n C 1/ holds in any case, this proves the inductive step, completing the
proof by induction.
So we have established that the jug problem has a preserved invariant, namely,
the amount of water in every jug is a linear combination of the capacities of the
jugs. Lemma 9.1.6 has an important corollary:
Corollary. In trying to get 4 gallons from 12- and 18-gallon jugs, and likewise to
get 32 gallons from 899- and 1147-gallon jugs,
Bruce will die!
Proof. By the Water Jugs Lemma 9.1.6, with 12- and 18-gallon jugs, the amount
in any jug is a linear combination of 12 and 18. This is always a multiple of 6 by
Lemma 9.1.2.2, so Bruce can’t get 4 gallons. Likewise, the amount in any jug using
899- and 1147-gallon jugs is a multiple of 31, so he can’t get 32 either.
But the Water Jugs Lemma doesn’t tell the complete story. For example, it leaves
open the question of getting 3 gallons into a jug using just 21- and 26-gallon jugs:
the only positive factor of both 21 and 26 is 1, and of course 1 divides 3, so the
Lemma neither rules out nor confirms the possibility of getting 3 gallons.
A bigger issue is that we’ve just managed to recast a pretty understandable question about water jugs into a technical question about linear combinations. This
might not seem like a lot of progress. Fortunately, linear combinations are closely
related to something more familiar, greatest common divisors, and will help us
solve the general water jug problem.
9.2
The Greatest Common Divisor
A common divisor of a and b is a number that divides them both. The greatest
common divisor of a and b is written gcd.a; b/. For example, gcd.18; 24/ D 6.
“mcs” — 2017/6/5 — 19:42 — page 307 — #315
9.2. The Greatest Common Divisor
307
As long as a and b are not both 0, they will have a gcd. The gcd turns out
to be very valuable for reasoning about the relationship between a and b and for
reasoning about integers in general. We’ll be making lots of use of gcd’s in what
follows.
Some immediate consequences of the definition of gcd are that
gcd.n; 1/ D 1
gcd.n; n/ D gcd.n; 0/ D jnj
f orn ¤ 0;
where the last equality follows from the fact that everything is a divisor of 0.
9.2.1
Euclid’s Algorithm
The first thing to figure out is how to find gcd’s. A good way called Euclid’s
algorithm has been known for several thousand years. It is based on the following
elementary observation.
Lemma 9.2.1. For b ¤ 0,
gcd.a; b/ D gcd.b; rem.a; b//:
Proof. By the Division Theorem 9.1.4,
a D qb C r
(9.2)
where r D rem.a; b/. So a is a linear combination of b and r, which implies that
any divisor of b and r is a divisor of a by Lemma 9.1.2.2. Likewise, r is a linear
combination a qb of a and b, so any divisor of a and b is a divisor of r. This
means that a and b have the same common divisors as b and r, and so they have
the same greatest common divisor.
Lemma 9.2.1 is useful for quickly computing the greatest common divisor of
two numbers. For example, we could compute the greatest common divisor of
1147 and 899 by repeatedly applying it:
gcd.1147; 899/ D gcd.899; rem.1147; 899//
„
ƒ‚
…
D248
D gcd .248; rem.899; 248/ D 155/
D gcd .155; rem.248; 155/ D 93/
D gcd .93; rem.155; 93/ D 62/
D gcd .62; rem.93; 62/ D 31/
D gcd .31; rem.62; 31/ D 0/
D 31
“mcs” — 2017/6/5 — 19:42 — page 308 — #316
308
Chapter 9
Number Theory
This calculation that gcd.1147; 899/ D 31 was how we figured out that with water
jugs of sizes 1147 and 899, Bruce dies trying to get 32 gallons.
On the other hand, applying Euclid’s algorithm to 26 and 21 gives
gcd.26; 21/ D gcd.21; 5/ D gcd.5; 1/ D 1;
so we can’t use the reasoning above to rule out Bruce getting 3 gallons into the big
jug. As a matter of fact, because the gcd here is 1, Bruce will be able to get any
number of gallons into the big jug up to its capacity. To explain this, we will need
a little more number theory.
Euclid’s Algorithm as a State Machine
Euclid’s algorithm can easily be formalized as a state machine. The set of states is
N2 and there is one transition rule:
.x; y/ ! .y; rem.x; y//;
(9.3)
for y > 0. By Lemma 9.2.1, the gcd stays the same from one state to the next. That
means the predicate
gcd.x; y/ D gcd.a; b/
is a preserved invariant on the states .x; y/. This preserved invariant is, of course,
true in the start state .a; b/. So by the Invariant Principle, if y ever becomes 0, the
invariant will be true and so
x D gcd.x; 0/ D gcd.a; b/:
Namely, the value of x will be the desired gcd.
What’s more x and therefore also y, gets to be 0 pretty fast. To see why, note
that starting from .x; y/, two transitions leads to a state whose the first coordinate
is rem.x; y/, which is at most half the size of x.4 Since x starts off equal to a and
gets halved or smaller every two steps, it will reach its minimum value—which is
gcd.a; b/—after at most 2 log a transitions. After that, the algorithm takes at most
one more transition to terminate. In other words, Euclid’s algorithm terminates
after at most 1 C 2 log a transitions.5
4 In
other words,
rem.x; y/ x=2
for 0 < y x:
(9.4)
This is immediate if y x=2, since the remainder of x divided by y is less than y by definition. On
the other hand, if y > x=2, then rem.x; y/ D x y < x=2.
5 A tighter analysis shows that at most log .a/ transitions are possible where ' is the golden ratio
'
p
.1 C 5/=2, see Problem 9.14.
“mcs” — 2017/6/5 — 19:42 — page 309 — #317
9.2. The Greatest Common Divisor
9.2.2
309
The Pulverizer
We will get a lot of mileage out of the following key fact:
Theorem 9.2.2. The greatest common divisor of a and b is a linear combination
of a and b. That is,
gcd.a; b/ D sa C t b;
for some integers s and t.6
We already know from Lemma 9.1.2.2 that every linear combination of a and b is
divisible by any common factor of a and b, so it is certainly divisible by the greatest
of these common divisors. Since any constant multiple of a linear combination is
also a linear combination, Theorem 9.2.2 implies that any multiple of the gcd is a
linear combination, giving:
Corollary 9.2.3. An integer is a linear combination of a and b iff it is a multiple of
gcd.a; b/.
We’ll prove Theorem 9.2.2 directly by explaining how to find s and t. This job
is tackled by a mathematical tool that dates back to sixth-century India, where it
was called kuttaka, which means “the Pulverizer.” Today, the Pulverizer is more
commonly known as the “Extended Euclidean Gcd Algorithm,” because it is so
close to Euclid’s algorithm.
For example, following Euclid’s algorithm, we can compute the gcd of 259
and 70 as follows:
gcd.259; 70/ D gcd.70; 49/
since rem.259; 70/ D 49
D gcd.49; 21/
since rem.70; 49/ D 21
D gcd.21; 7/
since rem.49; 21/ D 7
D gcd.7; 0/
since rem.21; 7/ D 0
D 7:
The Pulverizer goes through the same steps, but requires some extra bookkeeping
along the way: as we compute gcd.a; b/, we keep track of how to write each of
the remainders (49, 21, and 7, in the example) as a linear combination of a and b.
This is worthwhile, because our objective is to write the last nonzero remainder,
6 This result is often referred to as Bezout’s lemma, which is a misattribution since it was first
published in the West 150 years earlier by someone else, and was described a thousand years before
that by Indian mathematicians Aryabhata and Bhaskara.
“mcs” — 2017/6/5 — 19:42 — page 310 — #318
310
Chapter 9
Number Theory
which is the gcd, as such a linear combination. For our example, here is this extra
bookkeeping:
x
259
70
y
70
49
49
21
21
7
.rem.x; y// D x q y
49 D a 3 b
21 D b 1 49
D b 1 .a 3 b/
D
1aC4b
7 D 49 2 21
D .a 3 b/ 2 . 1 a C 4 b/
D 3 a 11 b
0
We began by initializing two variables, x D a and y D b. In the first two columns
above, we carried out Euclid’s algorithm. At each step, we computed rem.x; y/
which equals x qcnt.x; y/ y. Then, in this linear combination of x and y, we
replaced x and y by equivalent linear combinations of a and b, which we already
had computed. After simplifying, we were left with a linear combination of a and
b equal to rem.x; y/, as desired. The final solution is boxed.
This should make it pretty clear how and why the Pulverizer works. If you have
doubts, you may work through Problem 9.13, where the Pulverizer is formalized as
a state machine and then verified using an invariant that is an extension of the one
used for Euclid’s algorithm.
Since the Pulverizer requires only a little more computation than Euclid’s algorithm, you can “pulverize” very large numbers very quickly by using this algorithm.
As we will soon see, its speed makes the Pulverizer a very useful tool in the field
of cryptography.
Now we can restate the Water Jugs Lemma 9.1.6 in terms of the greatest common
divisor:
Corollary 9.2.4. Suppose that we have water jugs with capacities a and b. Then
the amount of water in each jug is always a multiple of gcd.a; b/.
For example, there is no way to form 4 gallons using 3- and 6-gallon jugs, because 4 is not a multiple of gcd.3; 6/ D 3.
9.2.3
One Solution for All Water Jug Problems
Corollary 9.2.3 says that 3 can be written as a linear combination of 21 and 26,
since 3 is a multiple of gcd.21; 26/ D 1. So the Pulverizer will give us integers s
and t such that
3 D s 21 C t 26
(9.5)
“mcs” — 2017/6/5 — 19:42 — page 311 — #319
9.2. The Greatest Common Divisor
311
The coefficient s could be either positive or negative. However, we can readily
transform this linear combination into an equivalent linear combination
3 D s 0 21 C t 0 26
(9.6)
where the coefficient s 0 is positive. The trick is to notice that if in equation (9.5) we
increase s by 26 and decrease t by 21, then the value of the expression s 21 C t 26
is unchanged overall. Thus, by repeatedly increasing the value of s (by 26 at a
time) and decreasing the value of t (by 21 at a time), we get a linear combination
s 0 21 C t 0 26 D 3 where the coefficient s 0 is positive. (Of course t 0 must then be
negative; otherwise, this expression would be much greater than 3.)
Now we can form 3 gallons using jugs with capacities 21 and 26: We simply
repeat the following steps s 0 times:
1. Fill the 21-gallon jug.
2. Pour all the water in the 21-gallon jug into the 26-gallon jug. If at any time
the 26-gallon jug becomes full, empty it out, and continue pouring the 21gallon jug into the 26-gallon jug.
At the end of this process, we must have emptied the 26-gallon jug exactly t 0
times. Here’s why: we’ve taken s 0 21 gallons of water from the fountain, and
we’ve poured out some multiple of 26 gallons. If we emptied fewer than t 0 times,
then by (9.6), the big jug would be left with at least 3 C 26 gallons, which is more
than it can hold; if we emptied it more times, the big jug would be left containing
at most 3 26 gallons, which is nonsense. But once we have emptied the 26-gallon
jug exactly t 0 times, equation (9.6) implies that there are exactly 3 gallons left.
Remarkably, we don’t even need to know the coefficients s 0 and t 0 in order to
use this strategy! Instead of repeating the outer loop s 0 times, we could just repeat
until we obtain 3 gallons, since that must happen eventually. Of course, we have to
keep track of the amounts in the two jugs so we know when we’re done. Here’s the
“mcs” — 2017/6/5 — 19:42 — page 312 — #320
312
Chapter 9
Number Theory
solution using this approach starting with empty jugs, that is, at .0; 0/:
fill 21
!
.21; 0/
pour 21 into 26
!
fill 21
pour 21 to 26
fill 21
pour 21 to 26
fill 21
pour 21 to 26
fill 21
pour 21 to 26
! .21; 21/
! .21; 16/
! .21; 11/
!
fill 21
!
.21; 6/
.21; 1/
!
!
!
!
fill 21
pour 21 to 26
fill 21
pour 21 to 26
fill 21
pour 21 to 26
! .21; 12/
!
fill 21
!
.21; 7/
.21; 2/
!
!
!
!
fill 21
pour 21 to 26
fill 21
pour 21 to 26
fill 21
pour 21 to 26
! .21; 13/
!
.21; 8/
.1; 26/
empty 26
pour 21 to 26
! .11; 0/
!
empty 26
!
.6; 0/
.1; 0/
! .0; 16/
! .0; 11/
!
pour 21 to 26
!
!
!
!
!
.0; 6/
.0; 1/
.0; 22/
.17; 26/
.12; 26/
.7; 26/
.2; 26/
empty 26
pour 21 to 26
empty 26
pour 21 to 26
empty 26
pour 21 to 26
! .17; 0/
! .12; 0/
!
empty 26
!
.7; 0/
.2; 0/
! .0; 17/
! .0; 12/
!
pour 21 to 26
!
!
pour 21 to 26
! .21; 18/
.6; 26/
pour 21 to 26
! .16; 0/
pour 21 to 26
fill 21
! .21; 23/
.11; 26/
empty 26
!
pour 21 to 26
! .21; 17/
.16; 26/
pour 21 to 26
pour 21 to 26
fill 21
! .21; 22/
.0; 21/
empty 26
.0; 7/
.0; 2/
.0; 23/
.18; 26/
.13; 26/
.8; 26/
.3; 26/
empty 26
pour 21 to 26
empty 26
pour 21 to 26
empty 26
pour 21 to 26
! .18; 0/
! .13; 0/
!
empty 26
!
.8; 0/
.3; 0/
! .0; 18/
! .0; 13/
!
pour 21 to 26
!
.0; 8/
.0; 3/
The same approach works regardless of the jug capacities and even regardless of
the amount we’re trying to produce! Simply repeat these two steps until the desired
amount of water is obtained:
1. Fill the smaller jug.
2. Pour all the water in the smaller jug into the larger jug. If at any time the
larger jug becomes full, empty it out, and continue pouring the smaller jug
into the larger jug.
By the same reasoning as before, this method eventually generates every multiple—
up to the size of the larger jug—of the greatest common divisor of the jug capacities,
all the quantities we can possibly produce. No ingenuity is needed at all!
So now we have the complete water jug story:
Theorem 9.2.5. Suppose that we have water jugs with capacities a and b. For
any c 2 Œ0::a, it is possible to get c gallons in the size a jug iff c is a multiple of
gcd.a; b/.
“mcs” — 2017/6/5 — 19:42 — page 313 — #321
9.3. Prime Mysteries
9.2.4
313
Properties of the Greatest Common Divisor
It can help to have some basic gcd facts on hand:
Lemma 9.2.6.
a) gcd.ka; kb/ D k gcd.a; b/ for all k > 0.
b) .d j a AND d j b/ IFF d j gcd.a; b/.
c) If gcd.a; b/ D 1 and gcd.a; c/ D 1, then gcd.a; bc/ D 1.
d) If a j bc and gcd.a; b/ D 1, then a j c.
Showing how all these facts follow from Theorem 9.2.2 that gcd is a linear combination is a good exercise (Problem 9.11).
These properties are also simple consequences of the fact that integers factor
into primes in a unique way (Theorem 9.4.1). But we’ll need some of these facts
to prove unique factorization in Section 9.4, so proving them by appeal to unique
factorization would be circular.
9.3
Prime Mysteries
Some of the greatest mysteries and insights in number theory concern properties of
prime numbers:
Definition 9.3.1. A prime is a number greater than 1 that is divisible only by itself
and 1. A number other than 0, 1, and 1 that is not a prime is called composite.7
Here are three famous mysteries:
Twin Prime Conjecture There are infinitely many primes p such that p C 2 is also
a prime.
In 1966, Chen showed that there are infinitely many primes p such that p C2
is the product of at most two primes. So the conjecture is known to be almost
true!
Conjectured Inefficiency of Factoring Given the product of two large primes n D
pq, there is no efficient procedure to recover the primes p and q. That is,
no polynomial time procedure (see Section 3.5) is guaranteed to find p and
7 So
0, 1, and 1 are the only integers that are neither prime nor composite.
“mcs” — 2017/6/5 — 19:42 — page 314 — #322
314
Chapter 9
Number Theory
q in a number of steps bounded by a polynomial in the length of the binary
representation of n (not n itself). The length of the binary representation at
most 1 C log2 n.
The best algorithm known is the “number field sieve,” which runs in time
proportional to:
1=3
2=3
e 1:9.ln n/ .ln ln n/ :
This number grows more rapidly than any polynomial in log n and is infeasible when n has 300 digits or more.
Efficient factoring is a mystery of particular importance in computer science,
as we’ll explain later in this chapter.
Goldbach’s Conjecture We’ve already mentioned Goldbach’s Conjecture 1.1.6 several times: every even integer greater than two is equal to the sum of two
primes. For example, 4 D 2 C 2, 6 D 3 C 3, 8 D 3 C 5, etc.
In 1939, Schnirelman proved that every even number can be written as the
sum of not more than 300,000 primes, which was a start. Today, we know
that every even number is the sum of at most 6 primes.
Primes show up erratically in the sequence of integers. In fact, their distribution
seems almost random:
2; 3; 5; 7; 11; 13; 17; 19; 23; 29; 31; 37; 41; 43; : : : :
One of the great insights about primes is that their density among the integers has
a precise limit. Namely, let .n/ denote the number of primes up to n:
Definition 9.3.2.
.n/ WWD jfp 2 Œ2::n j p is primegj:
For example, .1/ D 0; .2/ D 1 and .10/ D 4, because 2, 3, 5, and 7 are the
primes less than or equal to 10. Step by step, grows erratically according to the
erratic spacing between successive primes, but its overall growth rate is known to
smooth out to be the same as the growth of the function n= ln n:
Theorem 9.3.3 (Prime Number Theorem).
lim
n!1
.n/
D 1:
n= ln n
“mcs” — 2017/6/5 — 19:42 — page 315 — #323
9.4. The Fundamental Theorem of Arithmetic
315
Thus, primes gradually taper off. As a rule of thumb, about 1 integer out of every
ln n in the vicinity of n is a prime.
The Prime Number Theorem was conjectured by Legendre in 1798 and proved
a century later by de la Vallée Poussin and Hadamard in 1896. However, after his
death, a notebook of Gauss was found to contain the same conjecture, which he
apparently made in 1791 at age 15. (You have to feel sorry for all the otherwise
“great” mathematicians who had the misfortune of being contemporaries of Gauss.)
A proof of the Prime Number Theorem is beyond the scope of this text, but there
is a manageable proof (see Problem 9.22) of a related result that is sufficient for our
applications:
Theorem 9.3.4 (Chebyshev’s Theorem on Prime Density). For n > 1,
.n/ >
9.4
n
:
3 ln n
The Fundamental Theorem of Arithmetic
There is an important fact about primes that you probably already know: every
positive integer number has a unique prime factorization. So every positive integer
can be built up from primes in exactly one way. These quirky prime numbers are
the building blocks for the integers.
Since the value of a product of numbers is the same if the numbers appear in a
different order, there usually isn’t a unique way to express a number as a product
of primes. For example, there are three ways to write 12 as a product of primes:
12 D 2 2 3 D 2 3 2 D 3 2 2:
What’s unique about the prime factorization of 12 is that any product of primes
equal to 12 will have exactly one 3 and two 2’s. This means that if we sort the
primes by size, then the product really will be unique.
Let’s state this more carefully. A sequence of numbers is weakly decreasing
when each number in the sequence is at least as big as the numbers after it. Note
that a sequence of just one number as well as a sequence of no numbers—the empty
sequence—is weakly decreasing by this definition.
Theorem 9.4.1. [Fundamental Theorem of Arithmetic] Every positive integer is a
product of a unique weakly decreasing sequence of primes.
“mcs” — 2017/6/5 — 19:42 — page 316 — #324
316
Chapter 9
Number Theory
A Prime for Google
In late 2004 a billboard appeared in various locations around the country:
first 10-digit prime found
in consecutive digits of e
. com
Substituting the correct number for the expression in curly-braces produced the
URL for a Google employment page. The idea was that Google was interested in
hiring the sort of people that could and would solve such a problem.
How hard is this problem? Would you have to look through thousands or millions
or billions of digits of e to find a 10-digit prime? The rule of thumb derived from
the Prime Number Theorem says that among 10-digit numbers, about 1 in
ln 1010 23
is prime. This suggests that the problem isn’t really so hard! Sure enough, the
first 10-digit prime in consecutive digits of e appears quite early:
e D2:718281828459045235360287471352662497757247093699959574966
9676277240766303535475945713821785251664274274663919320030
599218174135966290435729003342952605956307381323286279434 : : :
“mcs” — 2017/6/5 — 19:42 — page 317 — #325
9.4. The Fundamental Theorem of Arithmetic
317
For example, 75237393 is the product of the weakly decreasing sequence of
primes
23; 17; 17; 11; 7; 7; 7; 3;
and no other weakly decreasing sequence of primes will give 75237393.8
Notice that the theorem would be false if 1 were considered a prime; for example,
15 could be written as 5 3, or 5 3 1, or 5 3 1 1, . . . .
There is a certain wonder in unique factorization, especially in view of the prime
number mysteries we’ve already mentioned. It’s a mistake to take it for granted,
even if you’ve known it since you were in a crib. In fact, unique factorization
actually fails for many
p integer-like sets of numbers, such as the complex numbers
of the form n C m
5 for m; n 2 Z (see Problem 9.25).
The Fundamental Theorem is also called the Unique Factorization Theorem,
which is a more descriptive and less pretentious, name—but we really want to get
your attention to the importance and non-obviousness of unique factorization.
9.4.1
Proving Unique Factorization
The Fundamental Theorem is not hard to prove, but we’ll need a couple of preliminary facts.
Lemma 9.4.2. If p is a prime and p j ab, then p j a or p j b.
Lemma 9.4.2 follows immediately from Unique Factorization: the primes in the
product ab are exactly the primes from a and from b. But proving the lemma this
way would be cheating: we’re going to need this lemma to prove Unique Factorization, so it would be circular to assume it. Instead, we’ll use the properties of gcd’s
and linear combinations to give an easy, noncircular way to prove Lemma 9.4.2.
Proof. One case is if gcd.a; p/ D p. Then the claim holds, because a is a multiple
of p.
Otherwise, gcd.a; p/ ¤ p. In this case gcd.a; p/ must be 1, since 1 and p are
the only positive divisors of p. Now gcd.a; p/ is a linear combination of a and p,
so we have 1 D sa C tp for some s; t. Then b D s.ab/ C .t b/p, that is, b is a
linear combination of ab and p. Since p divides both ab and p, it also divides their
linear combination b.
A routine induction argument extends this statement to:
8 The
“product” of just one number is defined to be that number, and the product of no numbers is
by convention defined to be 1. So each prime p is uniquely the product of the primes in the lengthone sequence consisting solely of p, and 1, which you will remember is not a prime, is uniquely the
product of the empty sequence.
“mcs” — 2017/6/5 — 19:42 — page 318 — #326
318
Chapter 9
Number Theory
Lemma 9.4.3. Let p be a prime. If p j a1 a2 an , then p divides some ai .
Now we’re ready to prove the Fundamental Theorem of Arithmetic.
Proof. Theorem 2.3.1 showed, using the Well Ordering Principle, that every positive integer can be expressed as a product of primes. So we just have to prove this
expression is unique. We will use Well Ordering to prove this too.
The proof is by contradiction: assume, contrary to the claim, that there exist
positive integers that can be written as products of primes in more than one way.
By the Well Ordering Principle, there is a smallest integer with this property. Call
this integer n, and let
n D p1 p2 pj ;
D q1 q2 qk ;
where both products are in weakly decreasing order and p1 q1 .
If q1 D p1 , then n=q1 would also be the product of different weakly decreasing
sequences of primes, namely,
p2 pj ;
q2 qk :
Since n=q1 < n, this can’t be true, so we conclude that p1 < q1 .
Since the pi ’s are weakly decreasing, all the pi ’s are less than q1 . But
q1 j n D p1 p2 pj ;
so Lemma 9.4.3 implies that q1 divides one of the pi ’s, which contradicts the fact
that q1 is bigger than all them.
9.5
Alan Turing
The man pictured in Figure 9.1 is Alan Turing, the most important figure in the
history of computer science. For decades, his fascinating life story was shrouded
by government secrecy, societal taboo, and even his own deceptions.
At age 24, Turing wrote a paper entitled On Computable Numbers, with an Application to the Entscheidungsproblem. The crux of the paper was an elegant way
to model a computer in mathematical terms. This was a breakthrough, because it
allowed the tools of mathematics to be brought to bear on questions of computation.
For example, with his model in hand, Turing immediately proved that there exist
“mcs” — 2017/6/5 — 19:42 — page 319 — #327
9.5. Alan Turing
319
Figure 9.1
Alan Turing
problems that no computer can solve—no matter how ingenious the programmer.
Turing’s paper is all the more remarkable because he wrote it in 1936, a full decade
before any electronic computer actually existed.
The word “Entscheidungsproblem” in the title refers to one of the 28 mathematical problems posed by David Hilbert in 1900 as challenges to mathematicians of
the 20th century. Turing knocked that one off in the same paper. And perhaps
you’ve heard of the “Church-Turing thesis”? Same paper. So Turing was a brilliant
guy who generated lots of amazing ideas. But this lecture is about one of Turing’s
less-amazing ideas. It involved codes. It involved number theory. And it was sort
of stupid.
Let’s look back to the fall of 1937. Nazi Germany was rearming under Adolf
Hitler, world-shattering war looked imminent, and—like us —Alan Turing was
pondering the usefulness of number theory. He foresaw that preserving military
secrets would be vital in the coming conflict and proposed a way to encrypt communications using number theory. This is an idea that has ricocheted up to our own
time. Today, number theory is the basis for numerous public-key cryptosystems,
digital signature schemes, cryptographic hash functions, and electronic payment
systems. Furthermore, military funding agencies are among the biggest investors
in cryptographic research. Sorry, Hardy!
Soon after devising his code, Turing disappeared from public view, and half a
century would pass before the world learned the full story of where he’d gone and
“mcs” — 2017/6/5 — 19:42 — page 320 — #328
320
Chapter 9
Number Theory
what he did there. We’ll come back to Turing’s life in a little while; for now, let’s
investigate the code Turing left behind. The details are uncertain, since he never
formally published the idea, so we’ll consider a couple of possibilities.
9.5.1
Turing’s Code (Version 1.0)
The first challenge is to translate a text message into an integer so we can perform
mathematical operations on it. This step is not intended to make a message harder
to read, so the details are not too important. Here is one approach: replace each
letter of the message with two digits (A D 01, B D 02, C D 03, etc.) and string all
the digits together to form one huge number. For example, the message “victory”
could be translated this way:
!
v
22
i
09
c
03
t
20
o
15
r
18
y
25
Turing’s code requires the message to be a prime number, so we may need to pad
the result with some more digits to make a prime. The Prime Number Theorem
indicates that padding with relatively few digits will work. In this case, appending
the digits 13 gives the number 2209032015182513, which is prime.
Here is how the encryption process works. In the description below, m is the
unencoded message (which we want to keep secret), m
b is the encrypted message
(which the Nazis may intercept), and k is the key.
Beforehand The sender and receiver agree on a secret key, which is a large prime k.
Encryption The sender encrypts the message m by computing:
m
b Dmk
Decryption The receiver decrypts m
b by computing:
m
b
D m:
k
For example, suppose that the secret key is the prime number k D 22801763489
and the message m is “victory.” Then the encrypted message is:
m
b Dmk
D 2209032015182513 22801763489
D 50369825549820718594667857
There are a couple of basic questions to ask about Turing’s code.
“mcs” — 2017/6/5 — 19:42 — page 321 — #329
9.5. Alan Turing
321
1. How can the sender and receiver ensure that m and k are prime numbers, as
required?
The general problem of determining whether a large number is prime or composite has been studied for centuries, and tests for primes that worked well
in practice were known even in Turing’s time. In the past few decades, very
fast primality tests have been found as described in the text box below.
Primality Testing
It’s easy
p to
˘ see that an integer n is prime iff it is not divisible by any number from
2 to
n (see Problem 1.14). Of course this naive way to test if n is prime takes
p
more than n steps, which is exponential in the size of n measured by the number
of digits in the decimal or binary representation of n. Through the early 1970’s,
no prime testing procedure was known that would never blow up like this.
In 1974, Volker Strassen invented a simple, fast probabilistic primality test.
Strassens’s test gives the right answer when applied to any prime number, but
has some probability of giving a wrong answer on a nonprime number. However,
the probability of a wrong answer on any given number is so tiny that relying on
the answer is the best bet you’ll ever make.
Still, the theoretical possibility of a wrong answer was intellectually
bothersome—even if the probability of being wrong was a lot less than the probability of an undetectable computer hardware error leading to a wrong answer.
Finally in 2002, in a breakthrough paper beginning with a quote from Gauss emphasizing the importance and antiquity of primality testing, Manindra Agrawal,
Neeraj Kayal, and Nitin Saxena presented an amazing, thirteen line description of
a polynomial time primality test.
This definitively places primality testing way below the exponential effort apparently needed for SAT and similar problems. The polynomial bound on the
Agrawal et al. test had degree 12, and subsequent research has reduced the degree to 5, but this is still too large to be practical, and probabilistic primality tests
remain the method used in practice today. It’s plausible that the degree bound can
be reduced a bit more, but matching the speed of the known probabilistic tests
remains a daunting challenge.
2. Is Turing’s code secure?
The Nazis see only the encrypted message m
b D m k, so recovering the
original message m requires factoring m
b. Despite immense efforts, no really
efficient factoring algorithm has ever been found. It appears to be a funda-
“mcs” — 2017/6/5 — 19:42 — page 322 — #330
322
Chapter 9
Number Theory
mentally difficult problem. So, although a breakthrough someday can’t be
ruled out, the conjecture that there is no efficient way to factor is widely
accepted. In effect, Turing’s code puts to practical use his discovery that
there are limits to the power of computation. Thus, provided m and k are
sufficiently large, the Nazis seem to be out of luck!
This all sounds promising, but there is a major flaw in Turing’s code.
9.5.2
Breaking Turing’s Code (Version 1.0)
Let’s consider what happens when the sender transmits a second message using
Turing’s code and the same key. This gives the Nazis two encrypted messages to
look at:
m
c1 D m1 k
and
m
c2 D m2 k
The greatest common divisor of the two encrypted messages, m
c1 and m
c2 , is the
secret key k. And, as we’ve seen, the gcd of two numbers can be computed very
efficiently. So after the second message is sent, the Nazis can recover the secret key
and read every message!
A mathematician as brilliant as Turing is not likely to have overlooked such a
glaring problem, and we can guess that he had a slightly different system in mind,
one based on modular arithmetic.
9.6
Modular Arithmetic
On the first page of his masterpiece on number theory, Disquisitiones Arithmeticae,
Gauss introduced the notion of “congruence.” Now, Gauss is another guy who
managed to cough up a half-decent idea every now and then, so let’s take a look
at this one. Gauss said that a is congruent to b modulo n iff n j .a b/. This is
written
a b .mod n/:
For example:
29 15
.mod 7/
because 7 j .29
15/:
It’s not useful to allow a moduli n 0, and so we will assume from now on that
moduli are positive.
There is a close connection between congruences and remainders:
Lemma 9.6.1 (Remainder).
ab
.mod n/ iff
rem.a; n/ D rem.b; n/:
“mcs” — 2017/6/5 — 19:42 — page 323 — #331
9.6. Modular Arithmetic
323
Proof. By the Division Theorem 9.1.4, there exist unique pairs of integers q1 ; r1
and q2 ; r2 such that:
a D q1 n C r1
b D q2 n C r2 ;
where r1 ; r2 2 Œ0::n/. Subtracting the second equation from the first gives:
a
b D .q1
q2 /n C .r1
r2 /;
where r1 r2 is in the interval . n; n/. Now a b .mod n/ if and only if n
divides the left-hand side of this equation. This is true if and only if n divides
the right-hand side, which holds if and only if r1 r2 is a multiple of n. But the
only multiple of n in . n; n/ is 0, so r1 r2 must in fact equal 0, that is, when
r1 WWD rem.a; n/ D r2 WWD rem.b; n/.
So we can also see that
29 15
.mod 7/
because rem.29; 7/ D 1 D rem.15; 7/:
Notice that even though “(mod 7)” appears on the end, the symbol isn’t any more
strongly associated with the 15 than with the 29. It would probably be clearer to
write 29 mod 7 15, for example, but the notation with the modulus at the end is
firmly entrenched, and we’ll just live with it.
The Remainder Lemma 9.6.1 explains why the congruence relation has properties like an equality relation. In particular, the following properties9 follow immediately:
Lemma 9.6.2.
aa
.mod n/
(reflexivity)
a b IFF b a
.mod n/
(symmetry)
.a b AND b c/ IMPLIES a c
.mod n/
(transitivity)
We’ll make frequent use of another immediate corollary of the Remainder Lemma 9.6.1:
Corollary 9.6.3.
a rem.a; n/
9 Binary
.mod n/
relations with these properties are called equivalence relations, see Section 10.10.
“mcs” — 2017/6/5 — 19:42 — page 324 — #332
324
Chapter 9
Number Theory
Still another way to think about congruence modulo n is that it defines a partition
of the integers into n sets so that congruent numbers are all in the same set. For
example, suppose that we’re working modulo 3. Then we can partition the integers
into 3 sets as follows:
f :::;
f :::;
f :::;
6;
5;
4;
3; 0; 3; 6; 9; : : :
2; 1; 4; 7; 10; : : :
1; 2; 5; 8; 11; : : :
g
g
g
according to whether their remainders on division by 3 are 0, 1, or 2. The upshot
is that when arithmetic is done modulo n, there are really only n different kinds
of numbers to worry about, because there are only n possible remainders. In this
sense, modular arithmetic is a simplification of ordinary arithmetic.
The next most useful fact about congruences is that they are preserved by addition and multiplication:
Lemma 9.6.4 (Congruence). If a b .mod n/ and c d .mod n/, then
aCc bCd
ac bd
.mod n/;
.mod n/:
(9.7)
(9.8)
Proof. Let’s start with 9.7. Since a b .mod n/, we have by definition that
n j .b a/ D .b C c/ .a C c/, so
aCc bCc
.mod n/:
Since c d .mod n/, the same reasoning leads to
bCc bCd
.mod n/:
Now transitivity (Lemma 9.6.2) gives
aCc bCd
.mod n/:
The proof for 9.8 is virtually identical, using the fact that if n divides .b
then it certainly also divides .bc ac/.
9.7
a/,
Remainder Arithmetic
The Congruence Lemma 9.6.1 says that two numbers are congruent iff their remainders are equal, so we can understand congruences by working out arithmetic with
remainders. And if all we want is the remainder modulo n of a series of additions,
multiplications, subtractions applied to some numbers, we can take remainders at
every step so that the entire computation only involves number in the range Œ0::n/.
“mcs” — 2017/6/5 — 19:42 — page 325 — #333
9.7. Remainder Arithmetic
325
General Principle of Remainder Arithmetic
To find the remainder on division by n of the result of a series of additions and
multiplications, applied to some integers
replace each integer operand by its remainder on division by n,
keep each result of an addition or multiplication in the range Œ0::n/ by immediately replacing any result outside that range by its remainder on division by n.
For example, suppose we want to find
rem..444273456789 C 155558585555 /4036666666 ; 36/:
(9.9)
This looks really daunting if you think about computing these large powers and
then taking remainders. For example, the decimal representation of 444273456789
has about 20 million digits, so we certainly don’t want to go that route. But remembering that integer exponents specify a series of multiplications, we follow the
General Principle and replace the numbers being multiplied by their remainders.
Since rem.44427; 36/ D 3; rem.15555858; 36/ D 6, and rem.403; 36/ D 7, we
find that (9.9) equals the remainder on division by 36 of
.33456789 C 65555 /76666666 :
(9.10)
That’s a little better, but 33456789 has about a million digits in its decimal representation, so we still don’t want to compute that. But let’s look at the remainders of
the first few powers of 3:
rem.3; 36/ D 3
rem.32 ; 36/ D 9
rem.33 ; 36/ D 27
rem.34 ; 36/ D 9:
We got a repeat of the second step, rem.32 ; 36/ after just two more steps. This
means means that starting at 32 , the sequence of remainders of successive powers
of 3 will keep repeating every 2 steps. So a product of an odd number of at least
three 3’s will have the same remainder on division by 36 as a product of just three
3’s. Therefore,
rem.33456789 ; 36/ D rem.33 ; 36/ D 27:
“mcs” — 2017/6/5 — 19:42 — page 326 — #334
326
Chapter 9
Number Theory
What a win!
Powers of 6 are even easier because rem.62 ; 36/ D 0, so 0’s keep repeating after
the second step. Powers of 7 repeat after six steps, but on the fifth step you get a 1,
that is rem.76 ; 36/ D 1, so (9.10) successively simplifies to be the remainders of
the following terms:
.33456789 C 65555 /76666666
.33 C 62 65553 /.76 /1111111
.33 C 0 65553 /11111111
D 27:
Notice that it would be a disastrous blunder to replace an exponent by its remainder. The general principle applies to numbers that are operands of plus and
times, whereas the exponent is a number that controls how many multiplications to
perform. Watch out for this.
9.7.1
The ring Zn
It’s time to be more precise about the general principle and why it works. To begin,
let’s introduce the notation Cn for doing an addition and then immediately taking
a remainder on division by n, as specified by the general principle; likewise for
multiplying:
i Cn j WWD rem.i C j; n/;
i n j WWD rem.ij; n/:
Now the General Principle is simply the repeated application of the following
lemma.
Lemma 9.7.1.
rem.i C j; n/ D rem.i; n/ Cn rem.j; n/;
(9.11)
rem.ij; n/ D rem.i; n/ n rem.j; n/:
(9.12)
Proof. By Corollary 9.6.3, i rem.i; n/ and j rem.j; n/, so by the Congruence Lemma 9.6.4
i C j rem.i; n/ C rem.j; n/
.mod n/:
By Corollary 9.6.3 again, the remainders on each side of this congruence are equal,
which immediately gives (9.11). An identical proof applies to (9.12).
“mcs” — 2017/6/5 — 19:42 — page 327 — #335
9.8. Turing’s Code (Version 2.0)
327
The set of integers in the range Œ0::n/ together with the operations Cn and n is
referred to as Zn , the ring of integers modulo n. As a consequence of Lemma 9.7.1,
the familiar rules of arithmetic hold in Zn , for example:
.i n j / n k D i n .j n k/:
These subscript-n’s on arithmetic operations really clog things up, so instead
we’ll just write “(Zn )” on the side to get a simpler looking equation:
.i j / k D i .j k/ .Zn /:
In particular, all of the following equalities10 are true in Zn :
.i j / k D i .j k/
.i C j / C k D i C .j C k/
(associativity of );
(associativity of C);
1k Dk
(identity for );
0Ck Dk
(identity for C);
k C . k/ D 0
(inverse for C);
i Cj Dj Ci
i .j C k/ D .i j / C .i k/
i j Dj i
(commutativity of C)
(distributivity);
(commutativity of )
Associativity implies the familiar fact that it’s safe to omit the parentheses in
products:
k1 k2 km
comes out the same in Zn no matter how it is parenthesized.
The overall theme is that remainder arithmetic is a lot like ordinary arithmetic.
But there are a couple of exceptions we’re about to examine.
9.8
Turing’s Code (Version 2.0)
In 1940, France had fallen before Hitler’s army, and Britain stood alone against
the Nazis in western Europe. British resistance depended on a steady flow of sup10 A set with addition and multiplication operations that satisfy these equalities is known as a
commutative ring. In addition to Zn , the integers, rationals, reals, and polynomials with integer
coefficients are all examples of commutative rings. On the other hand, the set fT; Fg of truth values
with OR for addition and AND for multiplication is not a commutative ring because it fails to satisfy
one of these equalities. The n n matrices of integers are not a commutative ring because they fail
to satisfy another one of these equalities.
“mcs” — 2017/6/5 — 19:42 — page 328 — #336
328
Chapter 9
Number Theory
plies brought across the north Atlantic from the United States by convoys of ships.
These convoys were engaged in a cat-and-mouse game with German “U-boats”
—submarines—which prowled the Atlantic, trying to sink supply ships and starve
Britain into submission. The outcome of this struggle pivoted on a balance of information: could the Germans locate convoys better than the Allies could locate
U-boats, or vice versa?
Germany lost.
A critical reason behind Germany’s loss was not made public until 1974: Germany’s naval code, Enigma, had been broken by the Polish Cipher Bureau,11 and
the secret had been turned over to the British a few weeks before the Nazi invasion
of Poland in 1939. Throughout much of the war, the Allies were able to route convoys around German submarines by listening in to German communications. The
British government didn’t explain how Enigma was broken until 1996. When the
story was finally released (by the US), it revealed that Alan Turing had joined the
secret British codebreaking effort at Bletchley Park in 1939, where he became the
lead developer of methods for rapid, bulk decryption of German Enigma messages.
Turing’s Enigma deciphering was an invaluable contribution to the Allied victory
over Hitler.
Governments are always tight-lipped about cryptography, but the half-century
of official silence about Turing’s role in breaking Enigma and saving Britain may
be related to some disturbing events after the war—more on that later. Let’s get
back to number theory and consider an alternative interpretation of Turing’s code.
Perhaps we had the basic idea right (multiply the message by the key), but erred in
using conventional arithmetic instead of modular arithmetic. Maybe this is what
Turing meant:
Beforehand The sender and receiver agree on a large number n, which may be
made public. (This will be the modulus for all our arithmetic.) As in Version
1.0, they also agree that some prime number k < n will be the secret key.
Encryption As in Version 1.0, the message m should be another prime in Œ0::n/.
The sender encrypts the message m to produce m
b by computing mk, but this
time modulo n:
m
b WWD m k .Zn /
(9.13)
Decryption (Uh-oh.)
The decryption step is a problem. We might hope to decrypt in the same way as
before by dividing the encrypted message m
b by the key k. The difficulty is that m
b
11 See
http://en.wikipedia.org/wiki/Polish Cipher Bureau.
“mcs” — 2017/6/5 — 19:42 — page 329 — #337
9.9. Multiplicative Inverses and Cancelling
329
is the remainder when mk is divided by n. So dividing m
b by k might not even give
us an integer!
This decoding difficulty can be overcome with a better understanding of when it
is ok to divide by k in modular arithmetic.
9.9
Multiplicative Inverses and Cancelling
The multiplicative inverse of a number x is another number x
x
1
1
such that
x D 1:
From now on, when we say “inverse,” we mean multiplicative (not relational) inverse.
For example, over the rational numbers, 1=3 is, of course, an inverse of 3, since,
1
3 D 1:
3
In fact, with the sole exception of 0, every rational number n=m has an inverse,
namely, m=n. On the other hand, over the integers, only 1 and -1 have inverses.
Over the ring Zn , things get a little more complicated. For example, 2 is a multiplicative inverse of 8 in Z15 , since
2 8 D 1 .Z15 /:
On the other hand, 3 does not have a multiplicative inverse in Z15 . We can prove
this by contradiction: suppose there was an inverse j for 3, that is
1 D 3 j .Z15 /:
Then multiplying both sides of this equality by 5 leads directly to the contradiction
5 D 0:
5 D 5 .3 j /
D .5 3/ j
D 0 j D 0 .Z15 /:
So there can’t be any such inverse j .
So some numbers have inverses modulo 15 and others don’t. This may seem a
little unsettling at first, but there’s a simple explanation of what’s going on.
“mcs” — 2017/6/5 — 19:42 — page 330 — #338
330
Chapter 9
9.9.1
Number Theory
Relative Primality
Integers that have no prime factor in common are called relatively prime.12 This
is the same as having no common divisor (prime or not) greater than 1. It’s also
equivalent to saying gcd.a; b/ D 1.
For example, 8 and 15 are relatively prime, since gcd.8; 15/ D 1. On the other
hand, 3 and 15 are not relatively prime, since gcd.3; 15/ D 3 ¤ 1. This turns out
to explain why 8 has an inverse over Z15 and 3 does not.
Lemma 9.9.1. If k 2 Œ0::n/ is relatively prime to n, then k has an inverse in Zn .13
Proof. If k is relatively prime to n, then gcd.n; k/ D 1 by definition of gcd. This
means we can use the Pulverizer from section 9.2.2 to find a linear combination of
n and k equal to 1:
sn C t k D 1:
So applying the General Principle of Remainder Arithmetic (Lemma 9.7.1), we get
.rem.s; n/ rem.n; n// C .rem.t; n/ rem.k; n// D 1 .Zn /:
But rem.n; n/ D 0, and rem.k; n/ D k since k 2 Œ0::n/, so we get
rem.t; n/ k D 1 .Zn /:
Thus, rem.t; n/ is a multiplicative inverse of k.
By the way, it’s nice to know that when they exist, inverses are unique. That is,
Lemma 9.9.2. If i and j are both inverses of k in Zn , then i D j .
Proof.
i D i 1 D i .k j / D .i k/ j D 1 j D j .Zn /:
So the proof of Lemma 9.9.1 shows that for any k relatively prime to n, the
inverse of k in Zn is simply the remainder of a coefficient we can easily find using
the Pulverizer.
Working with a prime modulus is attractive here because, like the rational and
real numbers, when p is prime, every nonzero number has an inverse in Zp . But
arithmetic modulo a composite is really only a little more painful than working
modulo a prime—though you may think this is like the doctor saying, “This is only
going to hurt a little,” before he jams a big needle in your arm.
12 Other
texts call them coprime.
works even in the extreme case that n D 1, because 0 1 .mod 1/, so it is consistent to
define 0 to be its own inverse in Z1 .
13 This
“mcs” — 2017/6/5 — 19:42 — page 331 — #339
9.9. Multiplicative Inverses and Cancelling
9.9.2
331
Cancellation
Another sense in which real numbers are nice is that it’s ok to cancel common
factors. In other words, if we know that t r D t s for real numbers r; s; t , then
as long as t ¤ 0, we can cancel the t’s and conclude that r D s. In general,
cancellation is not valid in Zn . For example,
3 10 D 3 5 .Z15 /;
(9.14)
but cancelling the 3’s leads to the absurd conclusion that 10 equals 5.
The fact that multiplicative terms cannot be cancelled is the most significant way
in which Zn arithmetic differs from ordinary integer arithmetic.
Definition 9.9.3. A number k is cancellable in Zn iff
ka Dkb
implies a D b .Zn /
for all a; b 2 Œ0::n/.
If a number is relatively prime to 15, it can be cancelled by multiplying by its
inverse. So cancelling works for numbers that have inverses:
Lemma 9.9.4. If k has an inverse in Zn , then it is cancellable.
But 3 is not relatively prime to 15, and that’s why it is not cancellable. More
generally, if k is not relatively prime to n, then we can show it isn’t cancellable in
Zn in the same way we showed that 3 is not cancellable in (9.14).
To summarize, we have
Theorem 9.9.5. The following are equivalent for k 2 Œ0::n/:14
gcd.k; n/ D 1;
k has an inverse in Zn ;
k is cancellable in Zn :
9.9.3
Decrypting (Version 2.0)
Multiplicative inverses are the key to decryption in Turing’s code. Specifically,
we can recover the original message by multiplying the encoded message by the
Zn -inverse j of the key:
m
b j D .m k/ j D m .k j / D m 1 D m .Zn /:
So all we need to decrypt the message is to find an inverse of the secret key k, which
will be easy using the Pulverizer—providing k has an inverse. But k is positive and
less than the modulus n, so one simple way to ensure that k is relatively prime to
the modulus is to have n be a prime number.
14 This
works even when n D 1—see the footnote following Lemma 9.9.1.
“mcs” — 2017/6/5 — 19:42 — page 332 — #340
332
Chapter 9
9.9.4
Number Theory
Breaking Turing’s Code (Version 2.0)
The Germans didn’t bother to encrypt their weather reports with the highly-secure
Enigma system. After all, so what if the Allies learned that there was rain off the
south coast of Iceland? But amazingly, this practice provided the British with a
critical edge in the Atlantic naval battle during 1941.
The problem was that some of those weather reports had originally been transmitted using Enigma from U-boats out in the Atlantic. Thus, the British obtained
both unencrypted reports and the same reports encrypted with Enigma. By comparing the two, the British were able to determine which key the Germans were
using that day and could read all other Enigma-encoded traffic. Today, this would
be called a known-plaintext attack.
Let’s see how a known-plaintext attack would work against Turing’s code. Suppose that the Nazis know both the plain text m and its
m
b D m k .Zn /;
and since m is positive and less than the prime n, the Nazis can use the Pulverizer
to find the Zn -inverse j of m. Now
j m
b D j .m k/ D .j m/ k D 1 k D k .Zn /:
So by computing j m
b D k .Zn /, the Nazis get the secret key and can then decrypt
any message!
This is a huge vulnerability, so Turing’s hypothetical Version 2.0 code has no
practical value. Fortunately, Turing got better at cryptography after devising this
code; his subsequent deciphering of Enigma messages surely saved thousands of
lives, if not the whole of Britain.
9.9.5
Turing Postscript
A few years after the war, Turing’s home was robbed. Detectives soon determined
that a former homosexual lover of Turing’s had conspired in the robbery. So they
arrested him—that is, they arrested Alan Turing—because at that time in Britain,
homosexuality was a crime punishable by up to two years in prison. Turing was
sentenced to a hormonal “treatment” for his homosexuality: he was given estrogen
injections. He began to develop breasts.
Three years later, Alan Turing, the founder of computer science, was dead. His
mother explained what happened in a biography of her own son. Despite her repeated warnings, Turing carried out chemistry experiments in his own home. Apparently, her worst fear was realized: by working with potassium cyanide while
eating an apple, he poisoned himself.
“mcs” — 2017/6/5 — 19:42 — page 333 — #341
9.10. Euler’s Theorem
333
However, Turing remained a puzzle to the very end. His mother was a devout
woman who considered suicide a sin. And, other biographers have pointed out,
Turing had previously discussed committing suicide by eating a poisoned apple.
Evidently, Alan Turing, who founded computer science and saved his country, took
his own life in the end, and in just such a way that his mother could believe it was
an accident.
Turing’s last project before he disappeared from public view in 1939 involved the
construction of an elaborate mechanical device to test a mathematical conjecture
called the Riemann Hypothesis. This conjecture first appeared in a sketchy paper by
Bernhard Riemann in 1859 and is now one of the most famous unsolved problems
in mathematics.
9.10
Euler’s Theorem
The RSA cryptosystem examined in the next section, and other current schemes
for encoding secret messages, involve computing remainders of numbers raised to
large powers. A basic fact about remainders of powers follows from a theorem due
to Euler about congruences.
Definition 9.10.1. For n > 0, define
.n/ WWD the number of integers in Œ0::n/, that are relatively prime to n.
This function is known as Euler’s function.15
For example, .7/ D 6 because all 6 positive numbers in Œ0::7/ are relatively
prime to the prime number 7. Only 0 is not relatively prime to 7. Also, .12/ D 4
since 1, 5, 7, and 11 are the only numbers in Œ0::12/ that are relatively prime to 12.16
More generally, if p is prime, then .p/ D p 1 since every positive number in
Œ0::p/ is relatively prime to p. When n is composite, however, the function gets
a little complicated. We’ll get back to it in the next section.
Euler’s Theorem is traditionally stated in terms of congruence:
Theorem (Euler’s Theorem). If n and k are relatively prime, then
k .n/ 1
.mod n/:
Things get simpler when we rephrase Euler’s Theorem in terms of Zn .
15 Some
16 Also,
texts call it Euler’s totient function.
.1/ D 1, but since we make no use of this fact, it only merits a footnote.
(9.15)
“mcs” — 2017/6/5 — 19:42 — page 334 — #342
334
Chapter 9
Number Theory
The Riemann Hypothesis
The formula for the sum of an infinite geometric series says:
1 C x C x2 C x3 C D
Substituting x D 21s , x D
sequence of equations:
1
3s ,
xD
1
5s ,
1
1
x
:
and so on for each prime number gives a
1
1
1
C 2s C 3s C D
s
2
2
2
1
1
1
1
1 C s C 2s C 3s C D
3
3
3
1
1
1
1
1 C s C 2s C 3s C D
5
5
5
1
::
:
1C
1
1=2s
1
1=3s
1
1=5s
Multiplying together all the left-hand sides and all the right-hand sides gives:
1
X
1
D
ns
nD1
Y
p2primes
1
1
1=p s
:
The sum on the left is obtained by multiplying out all the infinite series and applying the Fundamental Theorem of Arithmetic. For example, the term 1=300s
in the sum is obtained by multiplying 1=22s from the first equation by 1=3s in
the second and 1=52s in the third. Riemann noted that every prime appears in the
expression on the right. So he proposed to learn about the primes by studying
the equivalent, but simpler expression on the left. In particular, he regarded s as
a complex number and the left side as a function .s/. Riemann found that the
distribution of primes is related to values of s for which .s/ D 0, which led to
his famous conjecture:
Definition 9.9.6. The Riemann Hypothesis: Every nontrivial zero of the zeta
function .s/ lies on the line s D 1=2 C ci in the complex plane.
A proof would immediately imply, among other things, a strong form of the Prime
Number Theorem.
Researchers continue to work intensely to settle this conjecture, as they have for
over a century. It is another of the Millennium Problems whose solver will earn
$1,000,000 from the Clay Institute.
“mcs” — 2017/6/5 — 19:42 — page 335 — #343
9.10. Euler’s Theorem
335
Definition 9.10.2. Let Zn be the integers in Œ0::n/, that are relatively prime to n:17
Zn WWD fk 2 Œ0::n/ j gcd.k; n/ D 1g:
(9.16)
Consequently,
ˇ ˇ
.n/ D ˇZn ˇ :
Theorem 9.10.3 (Euler’s Theorem for Zn ). For all k 2 Zn ,
k .n/ D 1 .Zn /:
(9.17)
Theorem 9.10.3 will follow from two very easy lemmas.
Let’s start by observing that Zn is closed under multiplication in Zn :
Lemma 9.10.4. If j; k 2 Zn , then j n k 2 Zn .
There are lots of easy ways to prove this (see Problem 9.69).
Definition 9.10.5. For any element k and subset S of Zn , let
kS WWD fk n s j s 2 Sg:
Lemma 9.10.6. If k 2 Zn and S Zn , then
jkS j D jS j:
Proof. Since k 2 Zn , by Theorem 9.9.5 it is cancellable. Therefore,
Œks D k t .Zn /
implies s D t:
So multiplying by k in Zn maps all the elements of S to distinct elements of kS ,
which implies S and kS are the same size.
Corollary 9.10.7. If k 2 Zn , then
kZn D Zn :
Proof. A product of elements in Zn remains in Zn by Lemma 9.10.4. So if k 2 Zn ,
then kZn Zn . But by Lemma 9.10.6, kZn and Zn are the same size, so they must
be equal.
Now we can complete the proof of Euler’s Theorem 9.10.3 for Zn :
17 Some
other texts use the notation n for Zn .
“mcs” — 2017/6/5 — 19:42 — page 336 — #344
336
Chapter 9
Number Theory
Proof. Let
P WWD k1 k2 k.n/ .Zn /
be the product in Zn of all the numbers in Zn . Let
Q WWD .k k1 / .k k2 / .k k.n/ / .Zn /
for some k 2 Zn . Factoring out k’s immediately gives
Q D k .n/ P .Zn /:
But Q is the same as the product of the numbers in kZn , and kZn D Zn , so we
realize that Q is the product of the same numbers as P , just in a different order.
Altogether, we have
P D Q D k .n/ P .Zn /:
Furthermore, P 2 Zn by Lemma 9.10.4, and so it can be cancelled from both sides
of this equality, giving
1 D k .n/ .Zn /:
Euler’s theorem offers another way to find inverses modulo n: if k is relatively
prime to n, then k .n/ 1 is a Zn -inverse of k, and we can compute this power of
k efficiently using fast exponentiation. However, this approach requires computing
.n/. In the next section, we’ll show that computing .n/ is easy if we know the
prime factorization of n. But we know that finding the factors of n is generally hard
to do when n is large, and so the Pulverizer remains the best approach to computing
inverses modulo n.
Fermat’s Little Theorem
For the record, we mention a famous special case of Euler’s Theorem that was
known to Fermat a century earlier.
Corollary 9.10.8 (Fermat’s Little Theorem). Suppose p is a prime and k is not a
multiple of p. Then
k p 1 1 .mod p/:
9.10.1
Computing Euler’s Function
RSA works using arithmetic modulo the product of two large primes, so we begin
with an elementary explanation of how to compute .pq/ for primes p and q:
“mcs” — 2017/6/5 — 19:42 — page 337 — #345
9.10. Euler’s Theorem
337
Lemma 9.10.9.
.pq/ D .p
1/.q
1/
for primes p ¤ q.
Proof. Since p and q are prime, any number that is not relatively prime to pq must
be a multiple of p or a multiple of q. Among the pq numbers in Œ0::pq/, there are
precisely q multiples of p and p multiples of q. Since p and q are relatively prime,
the only number in Œ0::pq/ that is a multiple of both p and q is 0. Hence, there are
p C q 1 numbers in Œ0::pq/ that are not relatively prime to n. This means that
.pq/ D pq
D .p
.p C q
1/.q
1/
1/;
as claimed.18
The following theorem provides a way to calculate .n/ for arbitrary n.
Theorem 9.10.10.
(a) If p is a prime, then .p k / D p k
pk
1
for k 1.
(b) If a and b are relatively prime, then .ab/ D .a/.b/.
Here’s an example of using Theorem 9.10.10 to compute .300/:
.300/ D .22 3 52 /
D .22 / .3/ .52 /
D .2
2
1
2 /.3
1
0
(by Theorem 9.10.10.(b))
3 /.5
2
1
5 /
(by Theorem 9.10.10.(a))
D 80:
Note that Lemma 9.10.9 also follows as a special case of Theorem 9.10.10.(b),
since we know that .p/ D p 1 for any prime p.
To prove Theorem 9.10.10.(a), notice that every pth number among the p k numbers in Œ0::p k / is divisible by p, and only these are divisible by p. So 1=p of these
numbers are divisible by p and the remaining ones are not. That is,
.p k / D p k
.1=p/p k D p k
pk
1
:
We’ll leave a proof of Theorem 9.10.10.(b) to Problem 9.63.
As a consequence of Theorem 9.10.10, we have
18 This
proof previews a kind of counting argument that we will explore more fully in Part III.
“mcs” — 2017/6/5 — 19:42 — page 338 — #346
338
Chapter 9
Number Theory
Corollary 9.10.11. For any number n, if p1 , p2 , . . . , pj are the (distinct) prime
factors of n, then
1
1
1
.n/ D n 1
1
1
:
p1
p2
pj
We’ll give another proof of Corollary 9.10.11 based on rules for counting in
Section 15.9.5.
9.11
RSA Public Key Encryption
Turing’s code did not work as he hoped. However, his essential idea—using number theory as the basis for cryptography—succeeded spectacularly in the decades
after his death.
In 1977, Ronald Rivest, Adi Shamir, and Leonard Adleman at MIT proposed a
highly secure cryptosystem, called RSA, based on number theory. The purpose of
the RSA scheme is to transmit secret messages over public communication channels. As with Turing’s codes, the messages transmitted are nonnegative integers of
some fixed size.
Moreover, RSA has a major advantage over traditional codes: the sender and
receiver of an encrypted message need not meet beforehand to agree on a secret key.
Rather, the receiver has both a private key, which they guard closely, and a public
key, which they distribute as widely as possible. A sender wishing to transmit a
secret message to the receiver encrypts their message using the receiver’s widelydistributed public key. The receiver can then decrypt the received message using
their closely held private key. The use of such a public key cryptography system
allows you and Amazon, for example, to engage in a secure transaction without
meeting up beforehand in a dark alley to exchange a key.
Interestingly, RSA does not operate modulo a prime, as Turing’s hypothetical
Version 2.0 may have, but rather modulo the product of two large primes—typically
primes that are hundreds of digits long. Also, instead of encrypting by multiplication with a secret key, RSA exponentiates to a secret power—which is why Euler’s
Theorem is central to understanding RSA.
The scheme for RSA public key encryption appears in the box.
If the message m is relatively prime to n, then a simple application of Euler’s
Theorem implies that this way of decoding the encrypted message indeed reproduces the original unencrypted message. In fact, the decoding always works—even
in (the highly unlikely) case that m is not relatively prime to n. The details are
worked out in Problem 9.84.
“mcs” — 2017/6/5 — 19:42 — page 339 — #347
9.11. RSA Public Key Encryption
339
The RSA Cryptosystem
A Receiver who wants to be able to receive secret numerical messages creates a
private key, which they keep secret, and a public key, which they make publicly
available. Anyone with the public key can then be a Sender who can publicly
send secret messages to the Receiver—even if they have never communicated or
shared any information besides the public key.
Here is how they do it:
Beforehand The Receiver creates a public key and a private key as follows.
1. Generate two distinct primes, p and q. These are used to generate the
private key, and they must be kept hidden. (In current practice, p and
q are chosen to be hundreds of digits long.)
2. Let n WWD pq.
3. Select an integer e 2 Œ0::n/ such that gcd.e; .p 1/.q 1// D 1.
The public key is the pair .e; n/. This should be distributed widely.
4. Let the private key d 2 Œ0::n/ be the inverse of e in the ring
Z.p 1/.q 1/ . This private key can be found using the Pulverizer. The
private key d should be kept hidden!
Encoding To transmit a message m 2 Œ0::n/ to Receiver, a Sender uses the
public key to encrypt m into a numerical message
m
b WWD me .Zn /:
The Sender can then publicly transmit m
b to the Receiver.
Decoding The Receiver decrypts message m
b back to message m using the private key:
mDm
bd .Zn /:
“mcs” — 2017/6/5 — 19:42 — page 340 — #348
340
Chapter 9
Number Theory
Why is RSA thought to be secure? It would be easy to figure out the private
key d if you knew p and q—you could do it the same way the Receiver does using
the Pulverizer. But assuming the conjecture that it is hopelessly hard to factor a
number that is the product of two primes with hundreds of digits, an effort to factor
n is not going to break RSA.
Could there be another approach to reverse engineer the private key d from the
public key that did not involve factoring n? Not really. It turns out that given just
the private and the public keys, it is easy to factor n19 (a proof of this is sketched
in Problem 9.86). So if we are confident that factoring is hopelessly hard, then we
can be equally confident that finding the private key just from the public key will
be hopeless.
But even if we are confident that an RSA private key won’t be found, this doesn’t
rule out the possibility of decoding RSA messages in a way that sidesteps the private key. It is an important unproven conjecture in cryptography that any way of
cracking RSA—not just by finding the secret key—would imply the ability to factor. This would be a much stronger theoretical assurance of RSA security than is
presently known.
But the real reason for confidence is that RSA has withstood all attacks by the
world’s most sophisticated cryptographers for nearly 40 years. Despite decades of
these attacks, no significant weakness has been found. That’s why the mathematical, financial, and intelligence communities are betting the family jewels on the
security of RSA encryption.
You can hope that with more studying of number theory, you will be the first to
figure out how to do factoring quickly and, among other things, break RSA. But
be further warned that even Gauss worked on factoring for years without a lot to
show for his efforts—and if you do figure it out, you might wind up meeting some
humorless fellows working for a Federal agency in charge of security. . . .
9.12
What has SAT got to do with it?
So why does society, or at least everybody’s secret codes, fall apart if there is an
efficient test for satisfiability (SAT), as we claimed in Section 3.5? To explain this,
remember that RSA can be managed computationally because multiplication of two
primes is fast, but factoring a product of two primes seems to be overwhelmingly
demanding.
19 In practice, for this reason, the public and private keys should be randomly chosen so that neither
is “too small.”
“mcs” — 2017/6/5 — 19:42 — page 341 — #349
9.13. References
341
Let’s begin with the observation from Section 3.2 that a digital circuit can be
described by a bunch of propositional formulas of about the same total size as the
circuit. So testing circuits for satisfiability is equivalent to the SAT problem for
propositional formulas (see Problem 3.22).
Now designing digital multiplication circuits is completely routine. We can easily build a digital “product checker” circuit out of AND, OR, and NOT gates with 1
output wire and 4n digital input wires. The first n inputs are for the binary representation of an integer i , the next n inputs for the binary representation of an integer
j , and the remaining 2n inputs for the binary representation of an integer k. The
output of the circuit is 1 iff ij D k and i; j > 1. A straightforward design for such
a product checker uses proportional to n2 gates.
Now here’s how to factor any number m with a length 2n binary representation
using a SAT solver. First, fix the last 2n digital inputs—the ones for the binary
representation of k—so that k equals m.
Next, set the first of the n digital inputs for the representation of i to be 1. Do a
SAT test to see if there is a satisfying assignment of values for the remaining 2n 1
inputs used for the i and j representations. That is, see if the remaining inputs for
i and j can be filled in to cause the circuit to give output 1. If there is such an
assignment, fix the first i -input to be 1, otherwise fix it to be 0. So now we have set
the first i -input equal to the first digit of the binary representations of an i such that
ij D m.
Now do the same thing to fix the second of the n digital inputs for the representation of i , and then third, proceeding in this way through all the n inputs for the
number i . At this point, we have the complete n-bit binary representation of an
i > 1 such ij D m for some j > 1. In other words, we have found an integer i
that is a factor of m. We can now find j by dividing m by i .
So after n SAT tests, we have factored m. This means that if SAT for digital
circuits with 4n inputs and about n2 gates could be determined by a procedure
taking a number of steps bounded above by a degree d polynomial in n, then 2n
digit numbers can be factored in n times this many steps, that is, with a number of
steps bounded by a polynomial of degree d C 1 in n. So if SAT could be solved in
polynomial time, then so could factoring, and consequently RSA would be “easy”
to break.
9.13
References
[3], [45]
“mcs” — 2017/6/5 — 19:42 — page 342 — #350
342
Chapter 9
Number Theory
Problems for Section 9.1
Practice Problems
Problem 9.1.
Prove that a linear combination of linear combinations of integers a0 ; : : : ; an is a
linear combination of a0 ; : : : ; an .
Class Problems
Problem 9.2.
A number is perfect if it is equal to the sum of its positive divisors, other than itself.
For example, 6 is perfect, because 6 D 1 C 2 C 3. Similarly, 28 is perfect, because
28 D 1 C 2 C 4 C 7 C 14. Explain why 2k 1 .2k 1/ is perfect when 2k 1 is
prime.20
Problems for Section 9.2
Practice Problems
Problem 9.3.
Let
x WWD 21212121;
y WWD 12121212:
Use the Euclidean algorithm to find the GCD of x and y. Hint: Looks scary, but
it’s not.
Problem 9.4.
20 Euclid
proved this 2300 years ago.
About 250 years ago, Euler proved the
converse:
every even perfect number is of this form (for a simple proof see
http://primes.utm.edu/notes/proofs/EvenPerfect.html).
It is not known
if there are any odd perfect numbers at all. It is also not known if there are an infinite number of even
perfect numbers. One of the charms of number theory is that simple results like those given in this
problem lie at the brink of the unknown.
“mcs” — 2017/6/5 — 19:42 — page 343 — #351
9.13. References
343
Let
x WWD 1788 315 372 591000
22 /
y WWD 19.9
3712 533678 5929 :
(a) What is gcd.x; y/?
(b) What is lcm.x; y/? (“lcm” is least common multiple.)
Problem 9.5.
Prove that
gcd.a5 ; b 5 / D gcd.a; b/5
for all a; b 2 Z.
Problem 9.6.
Prove that gcd.m; n/ is the minimum positive value of any integer linear combination of integers m and n.
Class Problems
Problem 9.7.
Use the Euclidean Algorithm to prove that
gcd.13a C 8b; 5a C 3b/ D gcd.a; b/:
Problem 9.8.
(a) Use the Pulverizer to find integers x; y such that
x30 C y22 D gcd.30; 22/:
(b) Now find integers x 0 ; y 0 with 0 y 0 < 30 such that
x 0 30 C y 0 22 D gcd.30; 22/
Problem 9.9. (a) Use the Pulverizer to find gcd.84; 108/
“mcs” — 2017/6/5 — 19:42 — page 344 — #352
344
Chapter 9
Number Theory
(b) Find integers x, y with 0 y < 84 such that
x 84 C y 108 D gcd.84; 108/:
(c) Is there a multiplicative inverse of 84 in Z108 ? If not briefly explain why,
otherwise find it.
Problem 9.10.
Indicate true or false for the following statements about the greatest common
divisor, and provide counterexamples for those that are false.
(a) If gcd.a; b/ ¤ 1 and gcd.b; c/ ¤ 1, then gcd.a; c/ ¤ 1.
true
false
(b) If a j bc and gcd.a; b/ D 1, then a j c.
true
false
(c) gcd.an ; b n / D .gcd.a; b//n
true
false
(d) gcd.ab; ac/ D a gcd.b; c/.
true
false
(e) gcd.1 C a; 1 C b/ D 1 C gcd.a; b/.
true
false
(f) If an integer linear combination of a and b equals 1, then so does some integer
linear combination of a and b 2 .
true
false
(g) If no integer linear combination of a and b equals 2, then neither does any
integer linear combination of a2 and b 2 .
true
false
Problem 9.11.
For nonzero integers a, b, prove the following properties of divisibility and GCD’S.
You may use Theorem 9.2.2 that gcd.a; b/ is an integer linear combination of a and
b. You may not appeal to uniqueness of prime factorization Theorem 9.4.1, because
some of these properties are needed to prove unique factorization.)
(a) Every common divisor of a and b divides gcd.a; b/.
(b) gcd.ka; kb/ D k gcd.a; b/ for all k 2 N.
(c) If a j bc and gcd.a; b/ D 1, then a j c.
(d) If p j bc for some prime p then p j b or p j c.
(e) Let m be the smallest integer linear combination of a and b that is positive.
Show that m D gcd.a; b/.
“mcs” — 2017/6/5 — 19:42 — page 345 — #353
9.13. References
345
Homework Problems
Problem 9.12.
Here is a game you can analyze with number theory and always beat me. We start
with two distinct, positive integers written on a blackboard. Call them a and b.
Now we take turns. (I’ll let you decide who goes first.) On each turn, the player
must write a new positive integer on the board that is the difference of two numbers
that are already there. If a player cannot play, then they lose.
For example, suppose that 12 and 15 are on the board initially. Your first play
must be 3, which is 15 12. Then I might play 9, which is 12 3. Then you might
play 6, which is 15 9. Then I can’t play, so I lose.
(a) Show that every number on the board at the end of the game is a multiple of
gcd.a; b/.
(b) Show that every positive multiple of gcd.a; b/ up to max.a; b/ is on the board
at the end of the game.
(c) Describe a strategy that lets you win this game every time.
Problem 9.13.
Define the Pulverizer State machine to have:
states WWD N6
start state WWD .a; b; 0; 1; 1; 0/
(where a b > 0)
transitions WWD .x; y; s; t; u; v/ !
.y; rem.x; y/; u
sq; v
t q; s; t / (for q D qcnt.x; y/; y > 0):
(a) Show that the following properties are preserved invariants of the Pulverizer
machine:
gcd.x; y/ D gcd.a; b/;
(Inv1)
sa C t b D y; and
(Inv2)
ua C vb D x:
(Inv3)
(b) Conclude that the Pulverizer machine is partially correct.
(c) Explain why the machine terminates after at most the same number of transitions as the Euclidean algorithm.
“mcs” — 2017/6/5 — 19:42 — page 346 — #354
346
Chapter 9
Number Theory
Problem 9.14.
The Euclidean state machine is defined by the rule
.x; y/ ! .y; rem.x; y//;
(9.18)
for y > 0.
Prove that the smallest positive integers a b for which, starting in state .a; b/,
the state machine will make n transitions are F .n C 1/ and F .n/, where F .n/ is
the nth Fibonacci number.21
Hint: Induction.
Problem 9.15.
Let’s extend the jug filling scenario of Section 9.1.3 to three jugs and a receptacle.
Suppose the jugs can hold a, b and c gallons of water, respectively.
The receptacle can be used to store an unlimited amount of water, but has no
measurement markings. Excess water can be dumped into the drain. Among the
possible moves are:
1. fill a bucket from the hose,
2. pour from the receptacle to a bucket until the bucket is full or the receptacle
is empty, whichever happens first,
3. empty a bucket to the drain,
4. empty a bucket to the receptacle, and
5. pour from one bucket to another until either the first is empty or the second
is full.
(a) Model this scenario with a state machine. (What are the states? How does a
state change in response to a move?)
(b) Prove that Bruce can get k 2 N gallons of water into the receptacle using the
above operations if gcd.a; b; c/ j k.
p
5.25 shows that F .n/ ' n where ' is the golden ratio .1 C 5/=2. This implies
that the Euclidean algorithm halts after at most log' .a/ transitions, a somewhat smaller bound than
2 log2 a derived from equation (9.4).
21 Problem
“mcs” — 2017/6/5 — 19:42 — page 347 — #355
9.13. References
347
Problem 9.16.
The Binary GCD state machine computes the GCD of integers a; b > 0 using only
division by 2 and subtraction, which makes it run very efficiently on hardware that
uses binary representation of numbers. In practice, it runs more quickly than the
more famous Euclidean algorithm described in Section 9.2.1.
statesWWDN3
start stateWWD.a; b; 1/
transitionsWWD if min.x; y/ > 0; then .x; y; e/ !
(if 2 j x and 2 j y) (i1)
.x=2; y=2; 2e/
.x=2; y; e/
(else if 2 j x) (i2)
.x; y=2; e/
(else if 2 j y) (i3)
.x
y; y; e/
(else if x > y) (i4)
.y
x; x; e/
(else if y > x) (i5)
(otherwise (x D y)): (i6)
.1; 0; ex/
(a) Use the Invariant Principle to prove that if this machine stops, that is, reaches
a state .x; y; e/ in which no transition is possible, then e D gcd.a; b/.
(b) Prove that rule (i1)
.x; y; e/ ! .x=2; y=2; 2e/
is never executed after any of the other rules is executed.
(c) Prove that the machine reaches a final state in at most 1 C 3.log a C log b/
transitions. (This is a coarse bound; you may be able to get a better one.)
Problem 9.17.
Extend the binary gcd procedure of Problem 9.16 to obtain a new pulverizer that
uses only division by 2 and subtraction.
Hint: After the binary gcd procedure has factored out 2’s, it starts computing the
gcd.a; b/ for numbers a; b at least one of which is odd. It does this by successively
updating a pair of numbers .x; y/ such that gcd.x; y/ D gcd.a; b/. Extend the
procedure to find and update coefficients ux ; vx ; uy ; vy such that
ux a C vx b D x and uy a C vy b D y:
“mcs” — 2017/6/5 — 19:42 — page 348 — #356
348
Chapter 9
Number Theory
To see how to update the coefficients when at least one of a and b is odd and
ua C vb is even, show that either u and v are both even, or else u b and v C a
are both even.
Problem 9.18.
For any set A of integers,
gcd.A/ WWD the greatest common divisor of the elements of A.
The following useful property of gcd’s of sets is easy to take for granted:
Theorem.
gcd.A [ B/ D gcd.gcd.A/; gcd.B//;
(AuB)
for all finite sets A; B Z.
Theorem (AuB) has an easy proof as a Corollary of the Unique Factorization
Theorem. In this problem we develop a proof by induction just making repeated
use of Lemma 9.2.6.b :
.d j a AND d j b/ IFF d j gcd.a; b/:
(gcddiv)
The key to proving (AuB) will be generalizing (gcddiv) to finite sets.
Definition. For any subset A Z,
d j A WWD 8a 2 A: d j a:
(divdef)
Lemma.
d j A IFF d j gcd.A/:
(A-iff-gcdA)
for all d 2 Z and finite sets A Z.
(a) Prove that
gcd.a; gcd.b; c// D gcd.gcd.a; b/; c/
(gcd-assoc)
for all integers a; b; c.
From here on we write “a [ A” as an abbreviation for “fag [ A.”
(b) Prove that
d j .a [ b [ C / IFF d j .gcd.a; b/ [ C /
for all a; b; d 2 Z, and C Z.
(abCgcd)
“mcs” — 2017/6/5 — 19:42 — page 349 — #357
9.13. References
349
Proof.
d j .a [ b [ C / IFF .d j a/ AND .d j b/ AND .d j C /
IFF
.d j gcd.a; b// AND .d j C /
IFF
d j .gcd.a; b/ [ C /
(def (divdef) of divides)
by (gcddiv)
(def (divdef) of divides):
(c) Using parts (a) and (b), prove by induction on the size of A, that
d j .a [ A/
IFF
d j gcd.a; gcd.A//;
(divauA)
for all integers a; d and finite sets A Z. Explain why this proves (A-iff-gcdA).
(d) Prove Theorem (AuB).
(e) Conclude that gcd.A/ is an integer linear combination of the elements in A.
Exam Problems
Problem 9.19.
Prove that gcd.mb C r; b/ D gcd.b; r/ for all integers m; b; r.
Problem 9.20.
The Stata Center’s delicate balance depends on two buckets of water hidden in a
secret room. The big bucket has a volume of 25 gallons, and the little bucket has a
volume of 10 gallons. If at any time a bucket contains exactly 13 gallons, the Stata
Center will collapse. There is an interactive display where tourists can remotely
fill and empty the buckets according to certain rules. We represent the buckets as a
state machine.
The state of the machine is a pair .b; l/, where b is the volume of water in big
bucket, and l is the volume of water in little bucket.
(a) We informally describe some of the legal operations tourists can perform below. Represent each of the following operations as a transition of the state machine.
The first is done for you as an example.
1. Fill the big bucket.
.b; l/ ! .25; l/:
2. Empty the little bucket.
“mcs” — 2017/6/5 — 19:42 — page 350 — #358
350
Chapter 9
Number Theory
3. Pour the big bucket into the little bucket. You should have two cases defined
in terms of the state .b; l/: if all the water from the big bucket fits in the little
bucket, then pour all the water. If it doesn’t, pour until the little jar is full,
leaving some water remaining in the big jar.
(b) Use the Invariant Principle to show that, starting with empty buckets, the Stata
Center will never collapse. That is, the state .13; x/ in unreachable. (In verifying
your claim that the invariant is preserved, you may restrict to the representative
transitions of part (a).)
Problem 9.21.
Let
m D 29 524 74 117 ;
n D 23 722 11211 197 ;
p D 25 34 76042 1930 :
(a) What is the gcd.m; n; p/?
(b) What is the least common multiple lcm.m; n; p/?
Let k .n/ be the largest power of k that divides n, where k > 1. That is,
k .n/ WWD maxfi j k i divides ng:
If A is a nonempty set of nonnegative integers, define
k .A/ WWD fk .a/ j a 2 Ag:
(c) Express k .gcd.A// in terms of k .A/.
(d) Let p be a prime number. Express p .lcm.A// in terms of p .A/.
(e) Give an example of integers a; b where 6 .lcm.a; b// > max.6 .a/; 6 .b//.
Q
Q
(f) Let A be the product of all the elements in A. Express p .n/. A/ in terms
of p .A/.
(g) Let B also be a nonempty set of nonnegative integers. Conclude that
gcd.A [ B/ D gcd.gcd.A/; gcd.B//:
(9.19)
Hint: Consider p ./ of the left and right-hand sides of (9.19). You may assume
min.A [ B/ D min.min.A/; min.B//:
(9.20)
“mcs” — 2017/6/5 — 19:42 — page 351 — #359
9.13. References
351
Problems for Section 9.3
Homework Problems
Problem 9.22.
TBA: Chebyshvev lower bound in prime density, based on Shoup pp.75–76
Problems for Section 9.4
Practice Problems
Problem 9.23.
Let p be a prime number and a1 ; : : : ; an integers. Prove the following Lemma by
induction:
Lemma.
If p divides a product a1 a2 an ; then p divides some ai :
(*)
You may assume the case for n D 2 which was given by Lemma 9.4.2.
Be sure to clearly state and label your Induction Hypothesis, Base case(s), and
Induction step.
Class Problems
Problem 9.24. (a) Let m D 29 524 117 1712 and n D 23 722 11211 131 179 192 . What
is the gcd.m; n/? What is the least common multiple lcm.m; n/ of m and n? Verify
that
gcd.m; n/ lcm.m; n/ D mn:
(9.21)
(b) Describe in general how to find the gcd.m; n/ and lcm.m; n/ from the prime
factorizations of m and n. Conclude that equation (9.21) holds for all positive
integers m; n.
Homework Problems
Problem 9.25.
p
5 for some integers m; n
The set of complex
numbers that are equal topm C n
p
is called ZŒ
5. It will turn out that in ZŒ
5, not all numbers have unique
factorizations.
“mcs” — 2017/6/5 — 19:42 — page 352 — #360
352
Chapter 9
Number Theory
p
p
p
A sum or product of numbers in ZŒ
5 is in ZŒ
5, and since ZŒ
5 is a
subset of the complex numbers, all the usual rules for addition and multiplication
are true for it. But some weird things do happen. For example, the prime 29 has
factors:
p
(a) Find x; y 2 ZŒ
5 such that xy D 29 and x ¤ ˙1 ¤ y.
p
On the other hand, the
number
3
is
still
a
“prime”
even
in
ZŒ
5. More prep
p
5 is called irreducible over ZŒ
5 iff when xy D p
cisely, a number p p
2 ZŒ
for some x; y 2 ZŒ
5, either x D ˙1 or y D ˙1.
p
p
p
Claim. The numbers 3; 2 C
5, and 2
5 are irreducible over ZŒ
5.
In particular, this Claim implies that the number 9 factors into irreducibles over
p
ZŒ
5 in two different ways:
p
p
3 3 D 9 D .2 C
5/.2
5/:
p
So ZŒ
5 is an example of what is called a non-unique factorization domain.
To verify the Claim, we’ll appeal (without proof) to a familiar technical property
of complex numbers given in the following Lemma.
p
Definition. For p
a complex number c D r C si where r; s 2 R and i is
1, the
2
2
norm jcj of c is r C s .
Lemma. For c; d 2 C,
jcd j D jcj jd j :
p
(b) Prove that jxj2 ¤ 3 for all x 2 ZŒ
5.
p
(c) Prove that if x 2 ZŒ
5 and jxj D 1, then x D ˙1.
p
(d) Prove that if jxyj D 3 for some x; y 2 ZŒ
5, then x D ˙1 or y D ˙1.
p
2
Hint: jzj 2 N for z 2 ZŒ
5.
(e) Complete the proof of the Claim.
Problems for Section 9.6
Practice Problems
Problem 9.26.
Prove that if a b .mod 14/ and a b .mod 5/, then a b .mod 70/.
“mcs” — 2017/6/5 — 19:42 — page 353 — #361
9.13. References
353
Problem 9.27.
Show that there is an integer x such that
ax b
.mod n/
iff
gcd.a; n/ j b:
Class Problems
Problem 9.28. (a) Prove if n is not divisible by 3, then n2 1 .mod 3/.
(b) Show that if n is odd, then n2 1 .mod 8/.
(c) Conclude that if p is a prime greater than 3, then p 2
1 is divisible by 24.
Problem 9.29.
The values of polynomial p.n/ WWD n2 C n C 41 are prime for all the integers from
0 to 39 (see Section 1.1). Well, p didn’t work, but are there any other polynomials
whose values are always prime? No way! In fact, we’ll prove a much stronger
claim.
Definition. The set P of integer polynomial functions can be defined recursively:
Base cases:
the identity function IdZ .x/ WWD x is in P .
for any integer m the constant function cm .x/ WWD m is in P .
Constructor cases. If r; s 2 P , then r C s and r s 2 P .
(a) Using the recursive definition of integer polynomial functions given above,
prove by structural induction that for all q 2 P ,
j k
.mod n/
IMPLIES
q.j / q.k/
.mod n/;
for all integers j; k; n where n > 1.
Be sure to clearly state and label your Induction Hypothesis, Base case(s), and
Constructor step.
“mcs” — 2017/6/5 — 19:42 — page 354 — #362
354
Chapter 9
Number Theory
(b) We’ll say that q produces multiples if, for every integer greater than one in the
range of q, there are infinitely many different multiples of that integer in the range.
For example, if q.4/ D 7 and q produces multiples, then there are infinitely many
different multiples of 7 in the range of q, and of course, except for 7 itself, none of
these multiples is prime.
Prove that if q has positive degree and positive leading coefficient, then q produces
multiples. You may assume that every such polynomial is strictly increasing for
large arguments.
Part (b) implies that an integer polynomial with positive leading coefficient and
degree has infinitely many nonprimes in its range. This fact no longer holds true for
multivariate polynomials. An amazing consequence of Matiyasevich’s [35] solution to Hilbert’s Tenth Problem is that multivariate polynomials can be understood
as general purpose programs for generating sets of integers. If a set of nonnegative
integers can be generated by any program, then it equals the set of nonnegative
integers in the range of a multivariate integer polynomial! In particular, there is an
integer polynomial p.x1 ; : : : ; x7 / whose nonnegative values as x1 ; : : : ; x7 range
over N are precisely the set of all prime numbers!
Problems for Section 9.7
Practice Problems
Problem 9.30.
List the numbers of all statements below that are equivalent to
ab
.mod n/;
where n > 1 and a and b are integers. Briefly explain your reasoning.
i) 2a 2b .mod n/
ii) 2a 2b .mod 2n/
iii) a3 b 3 .mod n/
iv) rem.a; n/ D rem.b; n/
v) rem.n; a/ D rem.n; b/
vi) gcd.a; n/ D gcd.b; n/
“mcs” — 2017/6/5 — 19:42 — page 355 — #363
9.13. References
vii) gcd.n; a
viii) .a
355
b/ D n
b/ is a multiple of n
ix) 9k 2 Z: a D b C nk
Problem 9.31.
What is remainder.3101 ; 21/?
Homework Problems
Problem 9.32.
Prove that congruence is preserved by arithmetic expressions. Namely, prove that
ab
.mod n/;
(9.22)
then
eval.e; a/ eval.e; b/
.mod n/;
(9.23)
for all e 2 Aexp (see Section 7.4).
Problem 9.33.
A commutative ring is a set R of elements along with two binary operations ˚ and
˝ from R R to R. There is an element in R called the zero-element, 0, and
another element called the unit-element, 1. The operations in a commutative ring
satisfy the following ring axioms for r; s; t 2 R:
0
.r ˝ s/ ˝ t D r ˝ .s ˝ t /
(associativity of ˝);
.r ˚ s/ ˚ t D r ˚ .s ˚ t /
(associativity of ˚);
r ˚s Ds˚r
(commutativity of ˚)
r ˝s Ds˝r
(commutativity of ˝);
0˚r Dr
(identity for ˚);
1˝r Dr
(identity for ˝);
0
(inverse for ˚);
9r 2 R: r ˚ r D 0
r ˝ .s ˚ t / D .r ˝ s/ ˚ .r ˝ t /
(distributivity):
(a) Show that the zero-element is unique, that is, show that if z 2 R has the
property that
z ˚ r D r;
(9.24)
then z D 0.
“mcs” — 2017/6/5 — 19:42 — page 356 — #364
356
Chapter 9
Number Theory
(b) Show that additive inverses are unique, that is, show that
r ˚ r1 D 0 and
(9.25)
r ˚ r2 D 0
(9.26)
implies r1 D r2 .
(c) Show that multiplicative inverses are unique, that is, show that
r ˝ r1 D 1
r ˝ r2 D 1
implies r1 D r2 .
Problem 9.34.
This problem will use elementary properties of congruences to prove that every
positive integer divides infinitely many Fibonacci numbers.
A function f W N ! N that satisifies
f .n/ D c1 f .n
1/ C c2 f .n
2/ C C cd f .n
d/
(9.27)
for some ci 2 N and all n d is called a degree d linear-recursive.
A function f W N ! N has a degree d repeat modulo m at n and k when it
satisfies the following repeat congruences:
f .n
f .n
.d
f .n/ f .k/
1/ f .k 1/
::
:
1// f .k
.d
.mod m/;
.mod m/;
1//
.mod m/:
for k > n d 1.
For the rest of this problem, assume linear-recursive functions and repeats are
degree d > 0.
(a) Prove that if a linear-recursive function has a repeat modulo m at n and k, then
it has one at n C 1 and k C 1.
(b) Prove that for all m > 1, every linear-recursive function repeats modulo m at
n and k for some n; k 2 Œd 1::d C md /.
(c) A linear-recursive function is reverse-linear if its d th coefficient cd D ˙1.
Prove that if a reverse-linear function repeats modulo m at n and k for some n d ,
then it repeats modulo m at n 1 and k 1.
“mcs” — 2017/6/5 — 19:42 — page 357 — #365
9.13. References
357
(d) Conclude that every reverse-linear function must repeat modulo m at d
and .d 1/ C j for some j > 0.
1
(e) Conclude that if f is an reverse-linear function and f .k/ D 0 for some k 2
Œ0::d /, then every positive integer is a divisor of f .n/ for infinitely many n.
(f) Conclude that every positive integer is a divisor of infinitely many Fibonacci
numbers.
Hint: Start the Fibonacci sequence with the values 0,1 instead of 1, 1.
Class Problems
Problem 9.35.
Find
5555
remainder 98763456789 999
67893414259 ; 14 :
(9.28)
Problem 9.36.
The following properties of equivalence mod n follow directly from its definition
and simple properties of divisibility. See if you can prove them without looking up
the proofs in the text.
(a) If a b .mod n/, then ac bc .mod n/.
(b) If a b .mod n/ and b c .mod n/, then a c .mod n/.
(c) If a b .mod n/ and c d .mod n/, then ac bd .mod n/.
(d) rem.a; n/ a .mod n/.
Problem 9.37. (a) Why is a number written in decimal evenly divisible by 9 if and
only if the sum of its digits is a multiple of 9? Hint: 10 1 .mod 9/.
(b) Take a big number, such as 37273761261. Sum the digits, where every other
one is negated:
3 C . 7/ C 2 C . 7/ C 3 C . 7/ C 6 C . 1/ C 2 C . 6/ C 1 D
11
Explain why the original number is a multiple of 11 if and only if this sum is a
multiple of 11.
“mcs” — 2017/6/5 — 19:42 — page 358 — #366
358
Chapter 9
Number Theory
Problem 9.38.
At one time, the Guinness Book of World Records reported that the “greatest human
calculator” was a guy who could compute 13th roots of 100-digit numbers that were
13th powers. What a curious choice of tasks. . . .
In this problem, we prove
n13 n
.mod 10/
(9.29)
for all n.
(a) Explain why (9.29) does not follow immediately from Euler’s Theorem.
(b) Prove that
d 13 d
.mod 10/
(9.30)
for 0 d < 10.
(c) Now prove the congruence (9.29).
Problem 9.39. (a) Ten pirates find a chest filled with gold and silver coins. There
are twice as many silver coins in the chest as there are gold. They divide the gold
coins in such a way that the difference in the number of coins given to any two
pirates is not divisible by 10. They will only take the silver coins if it is possible
to divide them the same way. Is this possible, or will they have to leave the silver
behind? Prove your answer.
(b) There are also 3 sacks in the chest, containing 5, 49, and 51 rubies respectively. The treasurer of the pirate ship is bored and decides to play a game with the
following rules:
He can merge any two piles together into one pile, and
he can divide a pile with an even number of rubies into two piles of equal size.
He makes one move every day, and he will finish the game when he has divided the
rubies into 105 piles of one. Is it possible for him to finish the game?
Exam Problems
Problem 9.40.
Compute the remainder
rem.24989184637 6734598447 ; 15/;
carefully explaining the steps in your computation.
(rem15)
“mcs” — 2017/6/5 — 19:42 — page 359 — #367
9.13. References
359
Problem 9.41.
The sum of the digits of the base 10 representation of an integer is congruent modulo 9 to that integer. For example,
763 7 C 6 C 3
.mod 9/:
We can say that “9 is a good modulus for base 10.”
More generally, we’ll say “k is a good modulus for base b” when, for any nonnegative integer n, the sum of the digits of the base b representation of n is congruent to n modulo k. So 2 is not a good modulus for base 10 because
763 6 7 C 6 C 3
.mod 2/:
(a) What integers k > 1 are good moduli for base 10?
(b) Show that if b 1 .mod k/, then k is good for base b.
(c) Prove conversely, that if k is good for some base b 2, then b 1 .mod k/.
Hint: The base b representation of b.
(d) Exactly which integers k > 1 are good moduli for base 106?
Problem 9.42.
We define the sequence of numbers
(
1;
an D
an 1 C an 2 C an
3
C an
4;
for n 3,
for n > 3.
Use strong induction to prove that remainder.an ; 3/ D 1 for all n 0.
Problems for Section 9.8
Exam Problems
Problem 9.43.
Definition. The set P of single variable integer polynomials can be defined recursively:
Base cases:
“mcs” — 2017/6/5 — 19:42 — page 360 — #368
360
Chapter 9
Number Theory
the identity function, IdZ .x/ WWD x is in P .
for any integer m the constant function, cm .x/ WWD m is in P .
Constructor cases. If r; s 2 P , then r C s and r s 2 P .
Prove by structural induction that for all q 2 P ,
j k
.mod n/
IMPLIES
q.j / q.k/
.mod n/;
for all integers j; k; n where n > 1.
Be sure to clearly state and label your Induction Hypothesis, Base case(s), and
Constructor step.
Problems for Section 9.9
Practice Problems
Problem 9.44.
(a) Given inputs m; n 2 ZC , the Pulverizer will produce x; y 2 Z such that:
(b) Assume n > 1. Explain how to use the numbers x; y to find the inverse of m
modulo n when there is an inverse.
Problem 9.45.
What is the multiplicative inverse (mod 7) of 2? Reminder: by definition, your
answer must be an integer between 0 and 6.
Problem 9.46. (a) Find integer coefficients x, y such that 25xC32y D gcd.25; 32/.
(b) What is the inverse (mod 25) of 32?
Problem 9.47. (a) Use the Pulverizer to find integers s; t such that
40s C 7t D gcd.40; 7/:
(b) Adjust your answer to part (a) to find an inverse modulo 40 of 7 in Œ1::40/.
“mcs” — 2017/6/5 — 19:42 — page 361 — #369
9.13. References
361
Class Problems
Problem 9.48.
Two nonparallel lines in the real plane intersect at a point. Algebraically, this means
that the equations
y D m1 x C b1
y D m2 x C b2
have a unique solution .x; y/, provided m1 ¤ m2 . This statement would be false if
we restricted x and y to the integers, since the two lines could cross at a noninteger
point:
However, an analogous statement holds if we work over the integers modulo a
prime p. Find a solution to the congruences
y m1 x C b1
.mod p/
y m2 x C b2
.mod p/
when m1 6 m2 .mod p/. Express your solution in the form x ‹ .mod p/ and
y ‹ .mod p/ where the ‹’s denote expressions involving m1 , m2 , b1 and b2 . You
may find it helpful to solve the original equations over the reals first.
Problems for Section 9.10
Practice Problems
Problem 9.49.
Prove that k 2 Œ0::n/ has an inverse modulo n iff it has an inverse in Zn .
“mcs” — 2017/6/5 — 19:42 — page 362 — #370
362
Chapter 9
Number Theory
Problem 9.50.
What is rem.2479 ; 79/?
Hint: You should not need to do any actual multiplications!
Problem 9.51. (a) Prove that 2212001 has a multiplicative inverse modulo 175.
(b) What is the value of .175/, where is Euler’s function?
(c) What is the remainder of 2212001 divided by 175?
Problem 9.52.
How many numbers between 1 and 6042 (inclusive) are relatively prime to 3780?
Hint: 53 is a factor.
Problem 9.53.
How many numbers between 1 and 3780 (inclusive) are relatively prime to 3780?
Problem 9.54.
(a) What is the probability that an integer from 1 to 360 selected with uniform
probability is relatively prime to 360?
(b) What is the value of rem.798 ; 360/?
Class Problems
Problem 9.55.
Find the remainder of 261818181 divided by 297.
Hint: 1818181 D .180 10101/ C 1; use Euler’s theorem.
Problem 9.56.
77
Find the last digit of 77 .
“mcs” — 2017/6/5 — 19:42 — page 363 — #371
9.13. References
363
Problem 9.57.
Prove that n and n5 have the same last digit. For example:
25 D 32
795 D 3077056399
Problem 9.58.
Use Fermat’s theorem to find the inverse i of 13 modulo 23 with 1 i < 23.
Problem 9.59.
Let be Euler’s function.
(a) What is the value of .2/?
(b) What are three nonnegative integers k > 1 such that .k/ D 2?
(c) Prove that .k/ is even for k > 2.
Hint: Consider whether k has an odd prime factor or not.
(d) Briefly explain why .k/ D 2 for exactly three values of k.
Problem 9.60.
Suppose a; b are relatively prime and greater than 1. In this problem you will prove
the Chinese Remainder Theorem, which says that for all m; n, there is an x such
that
x m mod a;
(9.31)
x n mod b:
(9.32)
Moreover, x is unique up to congruence modulo ab, namely, if x 0 also satisfies (9.31) and (9.32), then
x 0 x mod ab:
(a) Prove that for any m; n, there is some x satisfying (9.31) and (9.32).
Hint: Let b 1 be an inverse of b modulo a and define ea WWD b
similarly. Let x D mea C neb .
1 b.
Define eb
(b) Prove that
Œx 0 mod a AND x 0 mod b
implies x 0 mod ab:
“mcs” — 2017/6/5 — 19:42 — page 364 — #372
364
Chapter 9
Number Theory
(c) Conclude that
x x 0 mod a AND x x 0 mod b
implies x x 0 mod ab:
(d) Conclude that the Chinese Remainder Theorem is true.
(e) What about the converse of the implication in part (c)?
Problem 9.61.
The order of k 2 Zn is the smallest positive m such that k m D 1 .Zn /.
(a) Prove that
k m D 1 .Zn / IMPLIES ord.k; n/ j m:
Hint: Take the remainder of m divided by the order.
Now suppose p > 2 is a prime of the form 2s C 1. For example, 21 C 1; 22 C
1; 24 C 1 are such primes.
(b) Conclude from part (a) that if 0 < k < p, then ord.k; p/ is a power of 2.
(c) Prove that ord.2; p/ D 2s and conclude that s is a power of 2.22
Hint: 2k
1 for k 2 Œ1::r is positive but too small to equal 0 .Zp /.
Homework Problems
Problem 9.62.
This problem is about finding square roots modulo a prime p.
(a) Prove that x 2 y 2 .mod p/ if and only if x y .mod p/ or x .mod p/. Hint: x 2 y 2 D .x C y/.x y/
y
An integer x is called a square root of n mod p when
x2 n
.mod p/:
An integer with a square root is called a square mod p. For example, if n is congruent to 0 or 1 mod p, then n is a square and it is it’s own square root.
So let’s assume that p is an odd prime and n 6 0 .mod p/. It turns out there is
a simple test we can perform to see if n is a square mod p:
k
of the form 22 C 1 are called Fermat numbers, so we can rephrase this conclusion as
saying that any prime of the form 2s C 1 must actually be a Fermat number. The Fermat numbers are
prime for k D 1; 2; 3; 4, but not for k D 5. In fact, it is not known if any Fermat number with k > 4
is prime.
22 Numbers
“mcs” — 2017/6/5 — 19:42 — page 365 — #373
9.13. References
365
Euler’s Criterion
i. If n is a square modulo p, then n.p
1/=2
ii. If n is not a square modulo p then n.p
1 .mod p/.
1/=2
1 .mod p/.
(b) Prove Case (i) of Euler’s Criterion. Hint: Use Fermat’s theorem.
(c) Prove Case (ii) of Euler’s Criterion. Hint: Use part (a)
(d) Suppose that p 3 .mod 4/, and n is a square mod p. Find a simple expression in terms of n and p for a square root of n. Hint: Write p as p D 4k C 3 and
use Euler’s Criterion. You might have to multiply two sides of an equation by n at
one point.
Problem 9.63.
Suppose a; b are relatively prime integers greater than 1. In this problem you will
prove that Euler’s function is multiplicative, that is, that
.ab/ D .a/.b/:
The proof is an easy consequence of the Chinese Remainder Theorem (Problem 9.60).
(a) Conclude from the Chinese Remainder Theorem that the function f W Œ0::ab/ !
Œ0::a/ Œ0::b/ defined by
f .x/ WWD .rem.x; a/; rem.x; b//
is a bijection.
(b) For any positive integer k let Zk be the integers in Œ0::k/ that are relatively
prime to k. Prove that the function f from part (a) also defines a bijection from
Zab to Za Zb .
(c) Conclude from the preceding parts of this problem that
.ab/ D .a/.b/:
(9.33)
(d) Prove Corollary 9.10.11: for any number n > 1, if p1 , p2 , . . . , pj are the
(distinct) prime factors of n, then
1
1
1
.n/ D n 1
1
1
:
p1
p2
pj
“mcs” — 2017/6/5 — 19:42 — page 366 — #374
366
Chapter 9
Number Theory
Problem 9.64.
Definition. Define the order of k over Zn to be
ord.k; n/ WWD minfm > 0 j k m D 1 .Zn /g:
If no positive power of k equals 1 in Zn , then ord.k; n/ WWD 1.
(a) Show that k 2 Zn iff k has finite order in Zn .
(b) Prove that for every k 2 Zn , the order of k over Zn divides .n/.
Hint: Let m D ord.k; n/. Consider the quotient and remainder of .n/ divided by
m.
Problem 9.65.
The general version of the Chinese Remainder Theorem (see Problem 9.60) extends
to more than two relatively prime moduli. Namely,
Theorem (General Chinese Remainder). Suppose a1 ; : : : ; ak are integers greater
than 1 and each is relatively prime to the others. Let n WWD a1 a2 ak . Then for
any integers m1 ; m2 ; : : : ; mk , there is a unique x 2 Œ0::n/ such that
x mi
.mod ai /;
for 1 i k.
The proof is a routine induction on k using a fact that follows immediately from
unique factorization: if a number is relatively prime to some other numbers, then it
is relatively prime to their product.
The General Chinese Remainder Theorem is the basis for an efficient approach
to performing a long series of additions and multiplications on “large” numbers.
Namely, suppose n was large, but each of the factors ai was small enough to be
handled by cheap and available arithmetic hardware units. Suppose a calculation
requiring many additions and multiplications needs to be performed. To do a single multiplication or addition of two large numbers x and y in the usual way in
this setting would involve breaking up the x and y into pieces small enough to be
handled by the arithmetic units, using the arithmetic units to perform additions and
multiplications on (many) pairs of small pieces, and then reassembling the pieces
into an answer. Moreover, the order in which these operations on pieces can be
performed is contrained by dependence among the pieces—because of “carries,”
“mcs” — 2017/6/5 — 19:42 — page 367 — #375
9.13. References
367
for example. And this process of breakup and reassembly has to be performed for
each addition and multiplication that needs to be performed on large numbers.
Explain how the General Chinese Remainder Theorem can be applied to perform a long series of additions and multiplications on “large” numbers much more
efficiently than the usual way described above.
Problem 9.66.
In this problem we’ll prove that for all integers a; m where m > 1,
am am
.m/
.mod m/:
(9.34)
Note that a and m need not be relatively prime.
Assume m D p1k1 pnkn for distinct primes, p1 ; : : : ; pn and positive integers
k1 ; : : : ; k n .
(a) Show that if pi does not divide a, then
a.m/ 1
.mod piki /:
(b) Show that if pi j a then
am
.m/
0
.mod piki /:
(9.35)
(c) Conclude (9.34) from the facts above.
Hint: am
am
.m/
D am
.m/ .a.m/
1/.
Problem 9.67.
The Generalized Postage Problem
Several other problems (2.7, 2.1, 5.32) work out which amounts of postage can
be formed using two stamps of given denominations. In this problem, we generalize
this to two stamps with arbitrary positive integer denominations a and b cents. Let’s
call an amount of postage that can be made from a and b cent stamps a makeable
amount.
Lemma. (Generalized Postage) If a and b are relatively prime positive integers,
then any integer greater than ab a b is makeable.
“mcs” — 2017/6/5 — 19:42 — page 368 — #376
368
Chapter 9
Number Theory
To prove the Lemma, consider the following array with a infinite rows:
0
b
2b
3b
::
:
.a
1/b .a
a
bCa
2b C a
3b C a
::
:
2a
b C 2a
2b C 2a
3b C 2a
::
:
1/b C a .a
1/b C 2a .a
3a
b C 3a
2b C 3a
3b C 3a
::
:
:::
:::
:::
:::
:::
1/b C 3a : : :
Note that every element in this array is clearly makeable.
(a) Suppose that n is at least as large as, and also congruent mod a to, the first
element in some row of this array. Explain why n must appear in the array.
(b) Prove that every integer from 0 to a
integers in the first column of this array.
1 is congruent modulo a to one of the
(c) Complete the proof of the Generalized Postage Lemma by using parts (a)
and (b) to conclude that every integer n > ab a b appears in the array, and
hence is makeable.
Hint: Suppose n is congruent mod a to the first element in some row. Assume n is
less than that element, and then show that n ab a b.
(d) (Optional) What’s more, ab
a
b is not makeable. Prove it.
(e) Explain why the following even more general lemma follows directly from the
Generalized Lemma and part (d).
Lemma. (Generalized2 Postage) If m and n are positive integers and gWWDgcd.m; n/ >
1, then with m and n cent stamps, you can only make amounts of postage that
are multiples of g. You can actually make any amount of postage greater than
.mn=g/ m n that is a multiple of g, but you cannot make .mn=g/ m n cents
postage.
(f) Optional and possibly unknown. Suppose you have three denominations of
stamps, a; b; c and gcd.a; b; c/ D 1. Give a formula for the smallest number nabc
such that you can make every amount of postage nabc .
Exam Problems
Problem 9.68.
What is the remainder of 639601 divided by 220?
“mcs” — 2017/6/5 — 19:42 — page 369 — #377
9.13. References
369
Problem 9.69.
Prove that if k1 and k2 are relatively prime to n, then so is k1 n k2 ,
(a) . . . using the fact that k is relatively prime to n iff k has an inverse modulo n.
Hint: Recall that k1 k2 k1 n k2 .mod n/.
(b) . . . using the fact that k is relatively prime to n iff k is cancellable modulo n.
(c) . . . using the Unique Factorization Theorem and the basic GCD properties such
as Lemma 9.2.1.
Problem 9.70.
Circle true or false for the statements below, and provide counterexamples for
those that are false. Variables, a; b; c; m; n range over the integers and m; n > 1.
(a) gcd.1 C a; 1 C b/ D 1 C gcd.a; b/.
true
false
true
false
(c) If a j bc and gcd.a; b/ D 1, then a j c.
true
false
(d) gcd.an ; b n / D .gcd.a; b//n
true
false
(e) If gcd.a; b/ ¤ 1 and gcd.b; c/ ¤ 1, then gcd.a; c/ ¤ 1.
true
false
true
false
true
false
true
false
true
false
true
false
(b) If a b .mod n/, then p.a/ p.b/ .mod n/
for any polynomial p.x/ with integer coefficients.
(f) If an integer linear combination of a and b equals 1,
then so does some integer linear combination of a2 and b 2 .
(g) If no integer linear combination of a and b equals 2,
then neither does any integer linear combination of a2 and b 2 .
(h) If ac bc .mod n/ and n does not divide c,
then a b .mod n/.
(i) Assuming a; b have inverses modulo n,
if a
1
b
1
.mod n/, then a b .mod n/.
(j) If ac bc .mod n/ and n does not divide c,
then a b .mod n/.
“mcs” — 2017/6/5 — 19:42 — page 370 — #378
370
Chapter 9
Number Theory
(k) If a b .mod .n// for a; b > 0, then c a c b .mod n/.
true
false
(l) If a b .mod nm/, then a b .mod n/.
true
false
Œa b .mod m/ AND a b .mod n/ iff Œa b .mod mn/
true
false
(n) If gcd.a; n/ D 1, then an
true
false
true
false
(m) If gcd.m; n/ D 1, then
1
1 .mod n/
(o) If a; b > 1, then
[a has a inverse mod b iff b has an inverse mod a].
Problem 9.71.
Parts (a) to (j) offer a collection of number theoretic assertions. Most of the
assertions require some side-conditions to be correct. A selection of possible sideconditions (1)–(8) are listed below.
Next to each of assertions, write True if the assertion is correct as stated. Write
False if none of the side-conditions ensure that the assertion is correct. Otherwise,
write the number of a side-condition that ensures the assertion will be true; to get
full credit your side-condition should not imply any other correct side-condition.
For example, you would only get part credit for choosing gcd.a; b/ > 2 as a sidecondition if only gcd.a; b/ > 1 was needed.
Assume all variables have appropriate integer values—for example, in the context of Zk , numbers are in Œ0::k/ and k > 1.
.i/: NOT.a divides b/ .ii/: gcd.a; b/ D 1
.iii/: gcd.a; b/ > 1
.iv/: gcd.a; b/ > 2
.v/: gcd.a; b/ D 1 D gcd.a; c/
.vi/: a is prime
.vii/: b is prime
.viii/: a; b both prime
(a) If a D b .Zn /, then p.a/ D p.b/ .Zn /,
for any polynomial p.x/ with integer coefficients.
(b) If a j bc, then a j c.
(c) gcd.1 C a; 1 C b/ D 1 C gcd.a; b/.
(d) No integer linear combination of a2 and b 2 equals one.
“mcs” — 2017/6/5 — 19:42 — page 371 — #379
9.13. References
371
(e) No integer linear combination of a2 and b 2 equals two.
(f) If ma D na .Zb / then m D n .Zb /.
(g) If b
1
Dc
1
.Za /, then b D c .Za /.
(h) If m D n .Z.a/ /, then b m D b n .Za /.
(i) If m D n .Zab /, then m D n .Zb /.
(j) ab D a .Zb /.
Problem 9.72.
Indicate whether the following statements are True or False. For each of the false
statements, give counterexamples. All variables range over the integers, Z.
(a) For all a and b, there are x and y such that: ax C by D 1.
(b) gcd.mb C r; b/ D gcd.r; b/ for all m; r and b.
(c) k p
1
1 .mod p/ for every prime p and every k.
(d) For primes p ¤ q, .pq/ D .p 1/.q 1/, where is Euler’s totient function.
(e) If a and b are relatively prime to d , then
Œac bc mod d 
IMPLIES
Œa b mod d :
Problem 9.73.
Find an integer k > 1 such that n and nk agree in their last three digits whenever n
is divisible by neither 2 nor 5. Hint: Euler’s theorem.
Problem 9.74.
(a) Explain why . 12/482 has a multiplicative inverse modulo 175.
(b) What is the value of .175/, where is Euler’s function?
“mcs” — 2017/6/5 — 19:42 — page 372 — #380
372
Chapter 9
Number Theory
(c) Call a number from 0 to 174 powerful iff some positive power of the number
is congruent to 1 modulo 175. What is the probability that a random number from
0 to 174 is powerful?
(d) What is the remainder of . 12/482 divided by 175?
Problem 9.75. (a) Calculate the remainder of 3586 divided by 29.
(b) Part (a) implies that the remainder of 3586 divided by 29 is not equal to 1. So
there there must be a mistake in the following proof, where all the congruences are
taken with modulus 29:
1 6 3586
(by part (a))
(9.36)
686
(since 35 6
.mod 29/)
(9.37)
28
(since 86 28
.mod 29/)
(9.38)
(by Fermat’s Little Theorem)
(9.39)
6
1
Identify the exact line containing the mistake and explain the logical error.
Problem 9.76. (a) Show that if p j n for some prime p and integer n > 0, then
.p 1/ j .n/.
(b) Conclude that .n/ is even for all n > 2.
Problem 9.77. (a) Calculate the value of .6042/.
Hint: 53 is a factor of 6042.
(b) Consider an integer k > 0 that is relatively prime to 6042. Explain why
k 9361 k .mod 6042/.
Hint: Use your solution to part (a).
“mcs” — 2017/6/5 — 19:42 — page 373 — #381
9.13. References
373
Problem 9.78.
Let
Sk D 1k C 2k C C p k ;
where p is an odd prime and k is a positive multiple of p
b 2 . p::0 such that
Sk a b .mod p/:
1. Find a 2 Œ0::p/ and
Problems for Section 9.11
Practice Problems
Problem 9.79.
Suppose a cracker knew how to factor the RSA modulus n into the product of
distinct primes p and q. Explain how the cracker could use the public key-pair
.e; n/ to find a private key-pair .d; n/ that would allow him to read any message
encrypted with the public key.
Problem 9.80.
Suppose the RSA modulus n D pq is the product of distinct 200 digit primes p and
q. A message m 2 Œ0::n/ is called dangerous if gcd.m; n/ D p, because such an m
can be used to factor n and so crack RSA. Circle the best estimate of the fraction
of messages in Œ0::n/ that are dangerous.
1
200
1
400
1
20010
1
10200
1
40010
1
10400
Problem 9.81.
Ben Bitdiddle decided to encrypt all his data using RSA. Unfortunately, he lost his
private key. He has been looking for it all night, and suddenly a genie emerges
from his lamp. He offers Ben a quantum computer that can perform exactly one
procedure on large numbers e; d; n. Which of the following procedures should Ben
choose to recover his data?
Find gcd.e; d /.
Find the prime factorization of n.
“mcs” — 2017/6/5 — 19:42 — page 374 — #382
374
Chapter 9
Number Theory
Determine whether n is prime.
Find rem.e d ; n/.
Find the inverse of e modulo n (the inverse of e in Zn /.
Find the inverse of e modulo .n/.
Class Problems
Problem 9.82.
Let’s try out RSA!
(a) Go through the beforehand steps.
Choose primes p and q to be relatively small, say in the range 10–40. In
practice, p and q might contain hundreds of digits, but small numbers are
easier to handle with pencil and paper.
Try e D 3; 5; 7; : : : until you find something that works. Use Euclid’s algorithm to compute the gcd.
Find d (using the Pulverizer).
When you’re done, put your public key on the board prominentally labelled “Public
Key.” This lets another team send you a message.
(b) Now send an encrypted message to another team using their public key. Select
your message m from the codebook below:
2 = Greetings and salutations!
3 = Yo, wassup?
4 = You guys are slow!
5 = All your base are belong to us.
6 = Someone on our team thinks someone on your team is kinda cute.
7 = You are the weakest link. Goodbye.
(c) Decrypt the message sent to you and verify that you received what the other
team sent!
Problem 9.83. (a) Just as RSA would be trivial to crack knowing the factorization
into two primes of n in the public key, explain why RSA would also be trivial to
crack knowing .n/.
“mcs” — 2017/6/5 — 19:42 — page 375 — #383
9.13. References
375
(b) Show that if you knew n, .n/, and that n was the product of two primes, then
you could easily factor n.
Problem 9.84.
A critical fact about RSA is, of course, that decrypting an encrypted message always gives back the original message m. Namely, if n D pq where p and q are
distinct primes, m 2 Œ0::pq/, and
d e 1
.mod .p
then
m
bd WWD me
d
1/.q
1//;
D m .Zn /:
(9.40)
We’ll now prove this.
(a) Explain why (9.40) follows very simply from Euler’s theorem when m is relatively prime to n.
All the rest of this problem is about removing the restriction that m be relatively
prime to n. That is, we aim to prove that equation (9.40) holds for all m 2 Œ0::n/.
It is important to realize that there is no practical reason to worry about—or to
bother to check for—this relative primality condition before sending a message m
using RSA. That’s because the whole RSA enterprise is predicated on the difficulty
of factoring. If an m ever came up that wasn’t relatively prime to n, then we could
factor n by computing gcd.m; n/. So believing in the security of RSA implies
believing that the liklihood of a message m turning up that was not relatively prime
to n is negligible.
But let’s be pure, impractical mathematicians and get rid of this technically unnecessary relative primality side condition, even if it is harmless. One gain for
doing this is that statements about RSA will be simpler without the side condition.
More important, the proof below illustrates a useful general method of proving
things about a number n by proving them separately for the prime factors of n.
(b) Prove that if p is prime and a 1 .mod p
1/, then
ma D m .Zp /:
(9.41)
(c) Give an elementary proof23 that if a b .mod pi / for distinct primes pi ,
then a b modulo the product of these primes.
(d) Note that (9.40) is a special case of
23 There
is no need to appeal to the Chinese Remainder Theorem.
“mcs” — 2017/6/5 — 19:42 — page 376 — #384
376
Chapter 9
Number Theory
Claim. If n is a product of distinct primes and a 1 .mod .n//, then
ma D m .Zn /:
Use the previous parts to prove the Claim.
Homework Problems
Problem 9.85.
Although RSA has successfully withstood cryptographic attacks for a more than a
quarter century, it is not known that breaking RSA would imply that factoring is
easy.
In this problem we will examine the Rabin cryptosystem that does have such
a security certification. Namely, if someone has the ability to break the Rabin
cryptosystem efficiently, then they also have the ability to factor numbers that are
products of two primes.
Why should that convince us that it is hard to break the cryptosystem efficiently?
Well, mathematicians have been trying to factor efficiently for centuries, and they
still haven’t figured out how to do it.
What is the Rabin cryptosystem? The public key will be a number N that is a
product of two very large primes p; q such that p q 3 .mod 4/. To send the
message m, send rem.m2 ; N /.24
The private key is the factorization of N , namely, the primes p; q. We need to
show that if the person being sent the message knows p; q, then they can decode
the message. On the other hand, if an eavesdropper who doesn’t know p; q listens
in, then we must show that they are very unlikely to figure out this message.
Say that s is a square modulo N if there is an m 2 Œ0; N / such that s m2
.mod N /. Such an m is a square root of s modulo N .
(a) What are the squares modulo 5? For each square in the interval Œ0; 5/, how
many square roots does it have?
(b) For each integer in Œ1::15/ that is relatively prime to 15, how many square
roots (modulo 15) does it have? Note that all the square roots are also relatively
prime to 15. We won’t go through why this is so here, but keep in mind that this is
a general phenomenon!
(c) Suppose that p is a prime such that p 3 .mod 4/. It turns out that squares
modulo p have exactly 2 square roots. First show that .p C 1/=4 is an integer.
24 We will see soon, that there are other numbers that would be encrypted by rem.m2 ; N /, so we’ll
have to disallow those other numbers as possible messages in order to make it possible to decode this
cryptosystem, but let’s ignore that for now.
“mcs” — 2017/6/5 — 19:42 — page 377 — #385
9.13. References
377
Next figure out the two square roots of 1 modulo p. Then show that you can find a
“square root mod a prime p” of a number by raising the number to the .p C 1/=4th
power. That is, given s, to find m such that s m2 .mod p/, you can compute
rem.s .pC1/=4 ; p/.
(d) The Chinese Remainder Theorem (Problem 9.60) implies that if p; q are distinct primes, then s is a square modulo pq if and only if s is a square modulo p and
s is a square modulo q. In particular, if s x 2 .x 0 /2 .mod p/ where x ¤ x 0 ,
and likewise s y 2 .y 0 /2 .mod q/ then s has exactly four square roots modulo
N , namely,
s .xy/2 .x 0 y/2 .xy 0 /2 .x 0 y 0 /2
.mod pq/:
So, if you know p; q, then using the solution to part (c), you can efficiently find the
square roots of s! Thus, given the private key, decoding is easy.
But what if you don’t know p; q?
Let’s assume that the evil message interceptor claims to have a program that can
find all four square roots of any number modulo N . Show that he can actually
use this program to efficiently find the factorization of N . Thus, unless this evil
message interceptor is extremely smart and has figured out something that the rest
of the scientific community has been working on for years, it is very unlikely that
this efficient square root program exists!
Hint: Pick r arbitrarily from Œ1::N /. If gcd.N; r/ > 1, then you are done (why?)
so you can halt. Otherwise, use the program to find all four square roots of r, call
them r; r; r 0 ; r 0 . Note that r 2 r 02 .mod N /. How can you use these roots to
factor N ?
(e) If the evil message interceptor knows that the message is the encoding one of
two possible candidate messages (that is, either “meet at dome at dusk” or “meet at
dome at dawn”) and is just trying to figure out which of the two, then can he break
this cryptosystem?
Problem 9.86.
You’ve seen how the RSA encryption scheme works, but why is it hard to break?
In this problem, you will see that finding private keys is as hard as finding the
prime factorizations of integers. Since there is a general consensus in the crypto
community (enough to persuade many large financial institutions, for example)
that factoring numbers with a few hundred digits requires astronomical computing
resources, we can therefore be sure it will take the same kind of overwhelming
“mcs” — 2017/6/5 — 19:42 — page 378 — #386
378
Chapter 9
Number Theory
effort to find RSA private keys of a few hundred digits. This means we can be
confident the private RSA keys are not somehow revealed by the public keys25 .
For this problem, assume that n D p q where p; q are both odd primes and that
e is the public key and d the private key of the RSA protocol.. Let c WWD e d 1.
(a) Show that .n/ divides c.
(b) Conclude that 4 divides c.
(c) Show that if gcd.r; n/ D 1, then r c 1 .mod n/:
A square root of m modulo n is an integer s 2 Œ0:n/ such that s 2 m .mod n/.
Here is a nice fact to know: when n is a product of two odd primes, then every
number m such that gcd.m; n/ D 1 has 4 square roots modulo n.
In particular, the number 1 has four square roots modulo n. The two trivial ones
are 1 and n 1 (which is 1 .mod n/). The other two are called the nontrivial
square roots of 1.
(d) Since you know c, then for any integer r you can also compute the remainder
y of r c=2 divided by n. So y 2 r c .mod n/. Now if r is relatively prime to n,
then y will be a square root of 1 modulo n by part (c).
Show that if y turns out to be a nontrivial root of 1 modulo n, then you can factor
n. Hint: From the fact that y 2 1 D .y C 1/.y 1/, show that y C 1 must be
divisible by exactly one of q and p.
(e) It turns out that at least half the positive integers r < n that are relatively
prime to n will yield y’s in part (d) that are nontrivial roots of 1. Conclude that if,
in addition to n and the public key e you also knew the private key d , then you can
be sure of being able to factor n.
Exam Problems
Problem 9.87.
Suppose Alice and Bob are using the RSA cryptosystem to send secure messages.
Each of them has a public key visible to everyone and a private key known only to
themselves, and using RSA in the usual way, they are able to send secret messages
to each other over public channels.
But a concern for Bob is how he knows that a message he gets is actually from
Alice—as opposed to some imposter claiming to be Alice. This concern can be met
by using RSA to add unforgeable “signatures” to messages. To send a message m
25 This
is a very weak kind of “security” property, because it doesn’t even rule out the possibility
of deciphering RSA encoded messages by some method that did not require knowing the private key.
Nevertheless, over twenty years experience supports the security of RSA in practice.
“mcs” — 2017/6/5 — 19:42 — page 379 — #387
9.13. References
379
to Bob with an unforgeable signature, Alice uses RSA encryption on her message
m, but instead using Bob’s public key to encrypt m, she uses her own private key
to obtain a message m1 . She then sends m1 as her “signed” message to Bob.
(a) Explain how Bob can read the original message m from Alice’s signed message m1 . (Let .nA ; eA / be Alice’s public key and dA her private key. Assume
m 2 Œ0::nA /.)
(b) Briefly explain why Bob can be confident, assuming RSA is secure, that m1
came from Alice rather than some imposter.
(c) Notice that not only Bob, but anyone can use Alice’s public key to reconstruct
her message m from its signed version m1 . So how can Alice send a secret signed
message to Bob over public channels?
“mcs” — 2017/6/5 — 19:42 — page 380 — #388
“mcs” — 2017/6/5 — 19:42 — page 381 — #389
10
Directed graphs & Partial Orders
Directed graphs, called digraphs for short, provide a handy way to represent how
things are connected together and how to get from one thing to another by following
those connections. They are usually pictured as a bunch of dots or circles with
arrows between some of the dots, as in Figure 10.1. The dots are called nodes or
vertices and the lines are called directed edges or arrows; the digraph in Figure 10.1
has 4 nodes and 6 directed edges.
Digraphs appear everywhere in computer science. For example, the digraph in
Figure 10.2 represents a communication net, a topic we’ll explore in depth in Chapter 11. Figure 10.2 has three “in” nodes (pictured as little squares) representing
locations where packets may arrive at the net, the three “out” nodes representing
destination locations for packets, and the remaining six nodes (pictured with little circles) represent switches. The 16 edges indicate paths that packets can take
through the router.
Another place digraphs emerge in computer science is in the hyperlink structure
of the World Wide Web. Letting the vertices x1 ; : : : ; xn correspond to web pages,
and using arrows to indicate when one page has a hyperlink to another, results in a
digraph like the one in Figure 10.3—although the graph of the real World Wide Web
would have n be a number in the billions and probably even the trillions. At first
glance, this graph wouldn’t seem to be very interesting. But in 1995, two students
at Stanford, Larry Page and Sergey Brin, ultimately became multibillionaires from
the realization of how useful the structure of this graph could be in building a search
engine. So pay attention to graph theory, and who knows what might happen!
b
a
c
d
Figure 10.1 A 4-node directed graph with 6 edges.
“mcs” — 2017/6/5 — 19:42 — page 382 — #390
382
Chapter 10
Directed graphs & Partial Orders
in0
in1
out0
in2
out1
out2
Figure 10.2 A 6-switch packet routing digraph.
x3
x4
x7
x2
x1
x6
Figure 10.3 Links among Web Pages.
x5
“mcs” — 2017/6/5 — 19:42 — page 383 — #391
10.1. Vertex Degrees
383
tail
head
e
u
v
Figure 10.4 A directed edge e D hu ! vi. The edge e starts at the tail vertex u
and ends at the head vertex v.
Definition 10.0.1. A directed graph G consists of a nonempty set V .G/, called the
vertices of G, and a set E.G/, called the edges of G. An element of V .G/ is called
a vertex. A vertex is also called a node; the words “vertex” and “node” are used
interchangeably. An element of E.G/ is called a directed edge. A directed edge is
also called an “arrow” or simply an “edge.” A directed edge starts at some vertex
u called the tail of the edge, and ends at some vertex v called the head of the edge,
as in Figure 10.4. Such an edge can be represented by the ordered pair .u; v/. The
notation hu ! vi denotes this edge.
There is nothing new in Definition 10.0.1 except for a lot of vocabulary. Formally, a digraph G is the same as a binary relation on the set, V D V .G/—that is,
a digraph is just a binary relation whose domain and codomain are the same set V .
In fact, we’ve already referred to the arrows in a relation G as the “graph” of G.
For example, the divisibility relation on the integers in the interval Œ1::12 could be
pictured by the digraph in Figure 10.5.
10.1
Vertex Degrees
The in-degree of a vertex in a digraph is the number of arrows coming into it, and
similarly its out-degree is the number of arrows out of it. More precisely,
Definition 10.1.1. If G is a digraph and v 2 V .G/, then
indeg.v/ WWD jfe 2 E.G/ j head.e/ D vgj
outdeg.v/ WWD jfe 2 E.G/ j tail.e/ D vgj
An immediate consequence of this definition is
Lemma 10.1.2.
X
indeg.v/ D
v2V .G/
Proof. Both sums are equal to jE.G/j.
X
outdeg.v/:
v2V .G/
“mcs” — 2017/6/5 — 19:42 — page 384 — #392
384
Chapter 10
Directed graphs & Partial Orders
4
2
8
10
5
12
6
1
7
3
9
11
Figure 10.5 The Digraph for Divisibility on f1; 2; : : : ; 12g.
10.2
Walks and Paths
Picturing digraphs with points and arrows makes it natural to talk about following
successive edges through the graph. For example, in the digraph of Figure 10.5,
you might start at vertex 1, successively follow the edges from vertex 1 to vertex
2, from 2 to 4, from 4 to 12, and then from 12 to 12 twice (or as many times as
you like). The sequence of edges followed in this way is called a walk through the
graph. A path is a walk which never visits a vertex more than once. So following
edges from 1 to 2 to 4 to 12 is a path, but it stops being a path if you go to 12 again.
The natural way to represent a walk is with the sequence of sucessive vertices it
went through, in this case:
1 2 4 12 12 12:
However, it is conventional to represent a walk by an alternating sequence of successive vertices and edges, so this walk would formally be
1 h1 ! 2i 2 h2 ! 4i 4 h4 ! 12i 12 h12 ! 12i 12 h12 ! 12i 12:
(10.1)
The redundancy of this definition is enough to make any computer scientist cringe,
but it does make it easy to talk about how many times vertices and edges occur on
the walk. Here is a formal definition:
Definition 10.2.1. A walk in a digraph is an alternating sequence of vertices and
edges that begins with a vertex, ends with a vertex, and such that for every edge
hu ! vi in the walk, vertex u is the element just before the edge, and vertex v is the
next element after the edge.
“mcs” — 2017/6/5 — 19:42 — page 385 — #393
10.2. Walks and Paths
385
So a walk v is a sequence of the form
v WWD v0 hv0 ! v1 i v1 hv1 ! v2 i v2 : : : hvk
1 ! vk i
vk
where hvi ! vi C1 i 2 E.G/ for i 2 Œ0::k/. The walk is said to start at v0 , to end at
vk , and the length jvj of the walk is defined to be k.
The walk is a path iff all the vi ’s are different, that is, if i ¤ j , then vi ¤ vj .
A closed walk is a walk that begins and ends at the same vertex. A cycle is a
positive length closed walk whose vertices are distinct except for the beginning and
end vertices.
Note that a single vertex counts as a length zero path that begins and ends at itself.
It also is a closed walk, but does not count as a cycle, since cycles by definition
must have positive length. Length one cycles are possible when a node has an
arrow leading back to itself. The graph in Figure 10.1 has none, but every vertex in
the divisibility relation digraph of Figure 10.5 is in a length one cycle. Length one
cycles are sometimes called self-loops.
Although a walk is officially an alternating sequence of vertices and edges, it
is completely determined just by the sequence of successive vertices on it, or by
the sequence of edges on it. We will describe walks in these ways whenever it’s
convenient. For example, for the graph in Figure 10.1,
.a; b; d /, or simply abd , is a (vertex-sequence description of a) length two
path,
.ha ! bi ; hb ! d i/, or simply ha ! bi hb ! d i, is (an edge-sequence description of) the same length two path,
abcbd is a length four walk,
dcbcbd is a length five closed walk,
bdcb is a length three cycle,
hb ! ci hc ! bi is a length two cycle, and
hc ! bi hb ai ha ! d i is not a walk. A walk is not allowed to follow edges
in the wrong direction.
If you walk for a while, stop for a rest at some vertex, and then continue walking,
you have broken a walk into two parts. For example, stopping to rest after following
two edges in the walk (10.1) through the divisibility graph breaks the walk into the
first part of the walk
1 h1 ! 2i 2 h2 ! 4i 4
(10.2)
“mcs” — 2017/6/5 — 19:42 — page 386 — #394
386
Chapter 10
Directed graphs & Partial Orders
from 1 to 4, and the rest of the walk
4 h4 ! 12i 12 h12 ! 12i 12 h12 ! 12i 12:
(10.3)
from 4 to 12, and we’ll say the whole walk (10.1) is the mergewalks (10.2) and (10.3).
In general, if a walk f ends with a vertex v and a walk r starts with the same vertex
v we’ll say that their merge f br is the walk that starts with f and continues with
r.1 Two walks can only be merged if the first walk ends at the same vertex v with
which the second one walk starts. Sometimes it’s useful to name the node v where
the walks merge; we’ll use the notation f b
v r to describe the merge of a walk f that
ends at v with a walk r that begins at v.
A consequence of this definition is that
Lemma 10.2.2.
jfbrj D jfj C jrj:
In the next section we’ll get mileage out of walking this way.
10.2.1
Finding a Path
If you were trying to walk somewhere quickly, you’d know you were in trouble if
you came to the same place twice. This is actually a basic theorem of graph theory.
Theorem 10.2.3. The shortest walk from one vertex to another is a path.
Proof. If there is a walk from vertex u to another vertex v ¤ u, then by the Well
Ordering Principle, there must be a minimum length walk w from u to v. We claim
w is a path.
To prove the claim, suppose to the contrary that w is not a path, meaning that
some vertex x occurs twice on this walk. That is,
w D eb
x fb
xg
for some walks e; f; g where the length of f is positive. But then “deleting” f yields
a strictly shorter walk
eb
xg
from u to v, contradicting the minimality of w.
Definition 10.2.4. The distance, dist .u; v/, in a graph from vertex u to vertex v is
the length of a shortest path from u to v.
1 It’s tempting to say the merge is the concatenation of the two walks, but that wouldn’t quite be
right because if the walks were concatenated, the vertex v would appear twice in a row where the
walks meet.
“mcs” — 2017/6/5 — 19:42 — page 387 — #395
10.3. Adjacency Matrices
387
As would be expected, this definition of distance satisfies:
Lemma 10.2.5. [The Triangle Inequality]
dist .u; v/ dist .u; x/ C dist .x; v/
for all vertices u; v; x with equality holding iff x is on a shortest path from u to v.
Of course, you might expect this property to be true, but distance has a technical
definition and its properties can’t be taken for granted. For example, unlike ordinary
distance in space, the distance from u to v is typically different from the distance
from v to u. So, let’s prove the Triangle Inequality
Proof. To prove the inequality, suppose f is a shortest path from u to x and r is
a shortest path from x to v. Then by Lemma 10.2.2, f b
x r is a walk of length
dist .u; x/ C dist .x; v/ from u to v, so this sum is an upper bound on the length of
the shortest path from u to v by Theorem 10.2.3.
Proof of the “iff” is in Problem 10.3.
Finally, the relationship between walks and paths extends to closed walks and
cycles:
Lemma 10.2.6. The shortest positive length closed walk through a vertex is a cycle
through that vertex.
The proof of Lemma 10.2.6 is essentially the same as for Theorem 10.2.3; see
Problem 10.4.
10.3
Adjacency Matrices
If a graph G has n vertices v0 ; v1 ; : : : ; vn 1 , a useful way to represent it is with an
n n matrix of zeroes and ones called its adjacency matrix AG . The ij th entry of
the adjacency matrix, .AG /ij , is 1 if there is an edge from vertex vi to vertex vj
and 0 otherwise. That is,
(
˝
˛
1 if vi ! vj 2 E.G/;
.AG /ij WWD
0 otherwise:
“mcs” — 2017/6/5 — 19:42 — page 388 — #396
388
Chapter 10
Directed graphs & Partial Orders
For example, let H be the 4-node graph shown in Figure 10.1. Its adjacency matrix
AH is the 4 4 matrix:
AH
a
a 0
D b 0
c 0
d 0
b
1
0
1
0
c
0
1
0
1
d
1
1
0
0
A payoff of this representation is that we can use matrix powers to count numbers
of walks between vertices. For example, there are two length two walks between
vertices a and c in the graph H :
a ha ! bi b hb ! ci c
a ha ! d i d hd ! ci c
and these are the only length two walks from a to c. Also, there is exactly one
length two walk from b to c and exactly one length two walk from c to c and from
d to b, and these are the only length two walks in H . It turns out we could have
read these counts from the entries in the matrix .AH /2 :
a
a 0
.AH /2 D b 0
c 0
d 0
b
0
1
0
1
c
2
1
1
0
d
1
0
1
0
More generally, the matrix .AG /k provides a count of the number of length k
walks between vertices in any digraph G as we’ll now explain.
Definition 10.3.1. The length-k walk counting matrix for an n-vertex graph G is
the n n matrix C such that
Cuv WWD the number of length-k walks from u to v:
(10.4)
Notice that the adjacency matrix AG is the length-1 walk counting matrix for G,
and that .AG /0 , which by convention is the identity matrix, is the length-0 walk
counting matrix.
Theorem 10.3.2. If C is the length-k walk counting matrix for a graph G, and D
is the length-m walk counting matrix, then CD is the length k C m walk counting
matrix for G.
“mcs” — 2017/6/5 — 19:42 — page 389 — #397
10.3. Adjacency Matrices
389
According to this theorem, the square .AG /2 of the adjacency matrix is the length
two walk counting matrix for G. Applying the theorem again to .AG /2 AG shows
that the length-3 walk counting matrix is .AG /3 . More generally, it follows by
induction that
Corollary 10.3.3. The length-k counting matrix of a digraph G is .AG /k , for all
k 2 N.
In other words, you can determine the number of length k walks between any
pair of vertices simply by computing the kth power of the adjacency matrix!
That may seem amazing, but the proof uncovers this simple relationship between
matrix multiplication and numbers of walks.
Proof of Theorem 10.3.2. Any length .kCm/ walk between vertices u and v begins
with a length k walk starting at u and ending at some vertex w followed by a length
m walk starting at w and ending at v. So the number of length .k C m/ walks from
u to v that go through w at the kth step equals the number Cuw of length k walks
from u to w, times the number Dw v of length m walks from w to v. We can get the
total number of length .k C m/ walks from u to v by summing, over all possible
vertices w, the number of such walks that go through w at the kth step. In other
words,
X
#length .k C m/ walks from u to v D
Cuw Dw v
(10.5)
w2V .G/
But the right-hand side of (10.5) is precisely the definition of .CD/uv . Thus, CD
is indeed the length-.k C m/ walk counting matrix.
10.3.1
Shortest Paths
The relation between powers of the adjacency matrix and numbers of walks is
cool—to us math nerds at least—but a much more important problem is finding
shortest paths between pairs of nodes. For example, when you drive home for
vacation, you generally want to take the shortest-time route.
One simple way to find the lengths of all the shortest paths in an n-vertex graph G
is to compute the successive powers of AG one by one up to the n 1st, watching
for the first power at which each entry becomes positive. That’s because Theorem 10.3.2 implies that the length of the shortest path, if any, between u and v,
that is, the distance from u to v, will be the smallest value k for which .AG /kuv is
nonzero, and if there is a shortest path, its length will be n 1. Refinements of
this idea lead to methods that find shortest paths in reasonably efficient ways. The
methods apply as well to weighted graphs, where edges are labelled with weights
or costs and the objective is to find least weight, cheapest paths. These refinements
“mcs” — 2017/6/5 — 19:42 — page 390 — #398
390
Chapter 10
Directed graphs & Partial Orders
are typically covered in introductory algorithm courses, and we won’t go into them
any further.
10.4
Walk Relations
A basic question about a digraph is whether there is a way to get from one particular
vertex to another. So for any digraph G we are interested in a binary relation G ,
called the walk relation on V .G/, where
u G v WWD there is a walk in G from u to v:
(10.6)
Similarly, there is a positive walk relation
u G C v WWD there is a positive length walk in G from u to v:
(10.7)
Definition 10.4.1. When there is a walk from vertex v to vertex w, we say that w
is reachable from v, or equivalently, that v is connected to w.
10.4.1
Composition of Relations
There is a simple way to extend composition of functions to composition of relations, and this gives another way to talk about walks and paths in digraphs.
Definition 10.4.2. Let R W B ! C and S W A ! B be binary relations. Then the
composition of R with S is the binary relation .R ı S / W A ! C defined by the
rule
a .R ı S / c WWD 9b 2 B: .a S b/ AND .b R c/:
(10.8)
This agrees with the Definition 4.3.1 of composition in the special case when R
and S are functions.2
Remembering that a digraph is a binary relation on its vertices, it makes sense to
compose a digraph G with itself. Then if we let G n denote the composition of G
with itself n times, it’s easy to check (see Problem 10.11) that G n is the length-n
walk relation:
a Gn b
iff
2 The reversal of the order of R
there is a length n walk in G from a to b:
and S in (10.8) is not a typo. This is so that relational composition
generalizes function composition. The value of function f composed with function g at an argument
x is f .g.x//. So in the composition f ı g, the function g is applied first.
“mcs” — 2017/6/5 — 19:42 — page 391 — #399
10.5. Directed Acyclic Graphs & Scheduling
391
This even works for n D 0, with the usual convention that G 0 is the identity relation
IdV .G/ on the set of vertices.3 Since there is a walk iff there is a path, and every
path is of length at most jV .G/j 1, we now have4
G D G 0 [ G 1 [ G 2 [ : : : [ G jV .G/j
1
D .G [ G 0 /jV .G/j
1
:
(10.9)
The final equality points to the use of repeated squaring as a way to compute G with log n rather than n 1 compositions of relations.
10.5
Directed Acyclic Graphs & Scheduling
Some of the prerequisites of MIT computer science subjects are shown in Figure 10.6. An edge going from subject s to subject t indicates that s is listed in the
catalogue as a direct prerequisite of t. Of course, before you can take subject t ,
you have to take not only subject s, but also all the prerequisites of s, and any prerequisites of those prerequisites, and so on. We can state this precisely in terms of
the positive walk relation: if D is the direct prerequisite relation on subjects, then
subject u has to be completed before taking subject v iff u D C v.
Of course it would take forever to graduate if this direct prerequisite graph had
a positive length closed walk. We need to forbid such closed walks, which by
Lemma 10.2.6 is the same as forbidding cycles. So, the direct prerequisite graph
among subjects had better be acyclic:
Definition 10.5.1. A directed acyclic graph (DAG) is a directed graph with no
cycles.
DAGs have particular importance in computer science. They capture key concepts used in analyzing task scheduling and concurrency control. When distributing
a program across multiple processors, we’re in trouble if one part of the program
needs an output that another part hasn’t generated yet! So let’s examine DAGs and
their connection to scheduling in more depth.
3 The
identity relation IdA on a set A is the equality relation:
a IdA b
iff a D b;
for a; b 2 A.
4 Equation (10.9) involves a harmless abuse of notation: we should have written
graph.G / D graph.G 0 / [ graph.G 1 / : : : :
“mcs” — 2017/6/5 — 19:42 — page 392 — #400
392
Chapter 10
Directed graphs & Partial Orders
New 6-3: SB in Computer Science and Engineering
Subjects
½+½
6.UAT
6.UAT
6.UAP
6.UAP
66 units
units
66 units
units
Advanced
Advanced Undergraduate
Undergraduate Subjects
Subjects
2
1
All subjects are 12 units
AUS
www.eecs.mit.edu/ug/newcurriculum/aus.html
http://
AUS http://www.eecs.mit.edu/ug/newcurriculum/aus.html
http://www.eecs.mit.edu/ug/newcurriculum/aus.html
Software
Software Lab
Lab
((http://www.eecs.mit.edu/ug/newcurriculum/verghese_6.005.html)
http://www.eecs.mit.edu/ug/newcurriculum/verghese_6.005.html)
3
Header
3
6.004
6.004
Foundation
comp
comp architecture
architecture
2
Introductory
(= 1 Institute Lab)
6.033
6.033
6.034
6.034
6.046
6.046
comp
comp sys
sys
AI
AI
adv
adv algorithms
algorithms
6.005*
6.005*
6.006*
6.006*
software
software
algorithms
algorithms
6.01*
6.01*
6.02*
6.02*
intro
intro EECS
EECS II
intro
intro EECS
EECS IIII
coreq
18.06 or 18.03
2
coreq
Math
(= 2 REST)
8.02
8.02
June 2009
*new subject
18.06
18.06
18.03
18.03
6.042
6.042
linear
linear algebra
algebra
diff
diff eqs
eqs
discrete
discrete math
math
Elementary
Elementary
exposure
exposure to
to programming
programming
(high
school,
(high school, IAP,
IAP, or
or 6.00)
6.00)
Figure 10.6 Subject prerequisites for MIT Computer Science (6-3) Majors.
“mcs” — 2017/6/5 — 19:42 — page 393 — #401
10.5. Directed Acyclic Graphs & Scheduling
left sock
393
right sock underwear
pants
left shoe
right shoe
shirt
tie
belt
jacket
Figure 10.7 DAG describing which clothing items have to be put on before others.
10.5.1
Scheduling
In a scheduling problem, there is a set of tasks, along with a set of constraints
specifying that starting certain tasks depends on other tasks being completed beforehand. We can map these sets to a digraph, with the tasks as the nodes and the
direct prerequisite constraints as the edges.
For example, the DAG in Figure 10.7 describes how a man might get dressed for
a formal occasion. As we describe above, vertices correspond to garments and the
edges specify which garments have to be put on before which others.
When faced with a set of prerequisites like this one, the most basic task is finding
an order in which to perform all the tasks, one at a time, while respecting the
dependency constraints. Ordering tasks in this way is known as topological sorting.
Definition 10.5.2. A topological sort of a finite DAG is a list of all the vertices
such that each vertex v appears earlier in the list than every other vertex reachable
from v.
There are many ways to get dressed one item at a time while obeying the constraints of Figure 10.7. We have listed two such topological sorts in Figure 10.8. In
“mcs” — 2017/6/5 — 19:42 — page 394 — #402
394
Chapter 10
Directed graphs & Partial Orders
Figure 10.8
ure 10.7
underwear
shirt
pants
belt
tie
jacket
left sock
right sock
left shoe
right shoe
left sock
shirt
tie
underwear
right sock
pants
right shoe
belt
jacket
left shoe
(a)
(b)
Two possible topological sorts of the prerequisites described in Fig.
fact, we can prove that every finite DAG has a topological sort. You can think of
this as a mathematical proof that you can indeed get dressed in the morning.
Topological sorts for finite DAGs are easy to construct by starting from minimal
elements:
Definition 10.5.3. An vertex v of a DAG D is minimum iff every other vertex is
reachable from v.
A vertex v is indexminimal vertex!directed graph—textbf minimal iff v is not
reachable from any other vertex.
It can seem peculiar to use the words “minimum” and “minimal” to talk about
vertices that start paths. These words come from the perspective that a vertex is
“smaller” than any other vertex it connects to. We’ll explore this way of thinking
about DAGs in the next section, but for now we’ll use these terms because they are
conventional.
One peculiarity of this terminology is that a DAG may have no minimum element
but lots of minimal elements. In particular, the clothing example has four minimal
elements: leftsock, rightsock, underwear, and shirt.
To build an order for getting dressed, we pick one of these minimal elements—
say, shirt. Now there is a new set of minimal elements; the three elements we didn’t
chose as step 1 are still minimal, and once we have removed shirt, tie becomes
minimal as well. We pick another minimal element, continuing in this way until all
elements have been picked. The sequence of elements in the order they were picked
will be a topological sort. This is how the topological sorts above were constructed.
“mcs” — 2017/6/5 — 19:42 — page 395 — #403
10.5. Directed Acyclic Graphs & Scheduling
395
So our construction shows:
Theorem 10.5.4. Every finite DAG has a topological sort.
There are many other ways of constructing topological sorts. For example, instead of starting from the minimal elements at the beginning of paths, we could
build a topological sort starting from maximal elements at the end of paths. In fact,
we could build a topological sort by picking vertices arbitrarily from a finite DAG
and simply inserting them into the list wherever they will fit.5
10.5.2
Parallel Task Scheduling
For task dependencies, topological sorting provides a way to execute tasks one after
another while respecting those dependencies. But what if we have the ability to
execute more than one task at the same time? For example, say tasks are programs,
the DAG indicates data dependence, and we have a parallel machine with lots of
processors instead of a sequential machine with only one. How should we schedule
the tasks? Our goal should be to minimize the total time to complete all the tasks.
For simplicity, let’s say all the tasks take the same amount of time and all the
processors are identical.
So given a finite set of tasks, how long does it take to do them all in an optimal
parallel schedule? We can use walk relations on acyclic graphs to analyze this
problem.
In the first unit of time, we should do all minimal items, so we would put on our
left sock, our right sock, our underwear, and our shirt.6 In the second unit of time,
we should put on our pants and our tie. Note that we cannot put on our left or right
shoe yet, since we have not yet put on our pants. In the third unit of time, we should
put on our left shoe, our right shoe, and our belt. Finally, in the last unit of time,
we can put on our jacket. This schedule is illustrated in Figure 10.9.
The total time to do these tasks is 4 units. We cannot do better than 4 units of
time because there is a sequence of 4 tasks that must each be done before the next.
We have to put on a shirt before pants, pants before a belt, and a belt before a jacket.
Such a sequence of items is known as a chain.
Definition 10.5.5. Two vertices in a DAG are comparable when one of them is
reachable from the other. A chain in a DAG is a set of vertices such that any two of
them are comparable. A vertex in a chain that is reachable from all other vertices
5 In
fact, the DAG doesn’t even need to be finite, but you’ll be relieved to know that we have no
need to go into this.
6 Yes, we know that you can’t actually put on both socks at once, but imagine you are being dressed
by a bunch of robot processors and you are in a big hurry. Still not working for you? Ok, forget about
the clothes and imagine they are programs with the precedence constraints shown in Figure 10.7.
“mcs” — 2017/6/5 — 19:42 — page 396 — #404
396
Chapter 10
Directed graphs & Partial Orders
A1
left sock
right sock underwear
pants
A2
A3
A4
left shoe
right shoe
shirt
tie
belt
jacket
Figure 10.9 A parallel schedule for the tasks-getting-dressed digraph in Figure 10.7. The tasks in Ai can be performed in step i for 1 i 4. A chain
of 4 tasks (the critical path in this example) is shown with bold edges.
“mcs” — 2017/6/5 — 19:42 — page 397 — #405
10.5. Directed Acyclic Graphs & Scheduling
397
in the chain is called a maximum element of the chain. A finite chain is said to end
at its maximum element.
The time it takes to schedule tasks, even with an unlimited number of processors,
is at least as large as the number of vertices in any chain. That’s because if we used
less time than the size of some chain, then two items from the chain would have to
be done at the same step, contradicting the precedence constraints. For this reason,
a largest chain is also known as a critical path. For example, Figure 10.9 shows
the critical path for the getting-dressed digraph.
In this example, we were able to schedule all the tasks with t steps, where t is
the size of the largest chain. A nice feature of DAGs is that this is always possible!
In other words, for any DAG, there is a legal parallel schedule that runs in t total
steps.
In general, a schedule for performing tasks specifies which tasks to do at successive steps. Every task a has to be scheduled at some step, and all the tasks that
have to be completed before task a must be scheduled for an earlier step. Here’s a
rigorous definition of schedule.
Definition 10.5.6. A partition of a set A is a set of nonempty subsets of A called
the blocks7 of the partition, such that every element of A is in exactly one block.
For example, one possible partition of the set fa; b; c; d; eg into three blocks is
fa; cg
fb; eg
fd g:
Definition 10.5.7. A parallel schedule for a DAG D is a partition of V .D/ into
blocks A0 ; A1 ; : : : ; such that when j < k, no vertex in Aj is reachable from any
vertex in Ak . The block Ak is called the set of elements scheduled at step k, and the
time of the schedule is the number of blocks. The maximum number of elements
scheduled at any step is called the number of processors required by the schedule.
A largest chain ending at an element a is called a critical path to a, and the
number of elements less than a in the chain is called the depth of a. So in any
possible parallel schedule, there must be at least depth .a/ steps before task a can
be started. In particular, the minimal elements are precisely the elements with depth
0.
There is a very simple schedule that completes every task in its minimum number of steps: just use a “greedy” strategy of performing tasks as soon as possible.
Schedule all the elements of depth k at step k. That’s how we found the above
schedule for getting dressed.
7 We
think it would be nicer to call them the parts of the partition, but “blocks” is the standard
terminology.
“mcs” — 2017/6/5 — 19:42 — page 398 — #406
398
Chapter 10
Directed graphs & Partial Orders
Theorem 10.5.8. A minimum time schedule for a finite DAG D consists of the sets
A0 ; A1 ; : : : ; where
Ak WWD fa 2 V .D/ j depth .a/ D kg:
We’ll leave to Problem 10.24 the proof that the sets Ak are a parallel schedule
according to Definition 10.5.7. We can summarize the story above in this way: with
an unlimited number of processors, the parallel time to complete all tasks is simply
the size of a critical path:
Corollary 10.5.9. Parallel time = size of critical path.
Things get more complex when the number of processors is bounded; see Problem 10.25 for an example.
10.5.3
Dilworth’s Lemma
Definition 10.5.10. An antichain in a DAG is a set of vertices such that no two elements in the set are comparable—no walk exists between any two different vertices
in the set.
Our conclusions about scheduling also tell us something about antichains.
Corollary 10.5.11. In a DAG D if the size of the largest chain is t , then V .D/ can
be partitioned into t antichains.
Proof. Let the antichains be the sets Ak WWD fa 2 V .D/ j depth .a/ D kg. It is an
easy exercise to verify that each Ak is an antichain (Problem 10.24).
Corollary 10.5.11 implies8 a famous result about acyclic digraphs:
Lemma 10.5.12 (Dilworth). For all t > 0, every DAG with n vertices must have
either a chain of size greater than t or an antichain of size at least n=t.
Proof. Assume that there is no chain of size greater than t . Let ` be the size of
the largest antichain. If we make a parallel schedule according to the proof of
Corollary 10.5.11, we create a number of antichains equal to the size of the largest
chain, which is less than or equal t . Each element belongs to exactly one antichain,
none of which are larger than `. So the total number of elements at most ` times
t—that is, `t n. Simple division implies that ` n=t.
8 Lemma 10.5.12 also follows from a more general result known as Dilworth’s Theorem, which
we will not discuss.
“mcs” — 2017/6/5 — 19:42 — page 399 — #407
10.6. Partial Orders
399
p
Corollary 10.5.13. Every DAG with n vertices has a chain of size greater than n
p
or an antichain of size at least n.
p
Proof. Set t D n in Lemma 10.5.12.
Example 10.5.14. When the man in our example is getting dressed, n D 10.
Try t D 3. There is a chain of size 4.
Try t D 4. There is no chain of size 5, but there is an antichain of size 4 10=4.
10.6
Partial Orders
After mapping the “direct prerequisite” relation onto a digraph, we were then able
to use the tools for understanding computer scientists’ graphs to make deductions
about something as mundane as getting dressed. This may or may not have impressed you, but we can do better. In the introduction to this chapter, we mentioned
a useful fact that bears repeating: any digraph is formally the same as a binary
relation whose domain and codomain are its vertices. This means that any binary
relation whose domain is the same as its codomain can be translated into a digraph!
Talking about the edges of a binary relation or the image of a set under a digraph
may seem odd at first, but doing so will allow us to draw important connections
between different types of relations. For instance, we can apply Dilworth’s lemma
to the “direct prerequisite” relation for getting dressed, because the graph of that
relation was a DAG.
But how can we tell if a binary relation is a DAG? And once we know that a
relation is a DAG, what exactly can we conclude? In this section, we will abstract
some of the properties that a binary relation might have, and use those properties
to define classes of relations. In particular, we’ll explain this section’s title, partial
orders.
10.6.1
The Properties of the Walk Relation in DAGs
To begin, let’s talk about some features common to all digraphs. Since merging a
walk from u to v with a walk from v to w gives a walk from u to w, both the walk
and positive walk relations have a relational property called transitivity:
Definition 10.6.1. A binary relation R on a set A is transitive iff
.a R b AND b R c/ IMPLIES a R c
for every a; b; c 2 A.
“mcs” — 2017/6/5 — 19:42 — page 400 — #408
400
Chapter 10
Directed graphs & Partial Orders
So we have
Lemma 10.6.2. For any digraph G the walk relations G C and G are transitive.
Since there is a length zero walk from any vertex to itself, the walk relation has
another relational property called reflexivity:
Definition 10.6.3. A binary relation R on a set A is reflexive iff a R a for all a 2 A.
Now we have
Lemma 10.6.4. For any digraph G, the walk relation G is reflexive.
We know that a digraph is a DAG iff it has no positive length closed walks. Since
any vertex on a closed walk can serve as the beginning and end of the walk, saying
a graph is a DAG is the same as saying that there is no positive length path from
any vertex back to itself. This means that the positive walk relation of D C of a
DAG has a relational property called irreflexivity.
Definition 10.6.5. A binary relation R on a set A is irreflexive iff
NOT .a
R a/
for all a 2 A.
So we have
Lemma 10.6.6. R is a DAG iff RC is irreflexive.
10.6.2
Strict Partial Orders
Here is where we begin to define interesting classes of relations:
Definition 10.6.7. A relation that is transitive and irreflexive is called a strict partial order.
A simple connection between strict partial orders and DAGs now follows from
Lemma 10.6.6:
Theorem 10.6.8. A relation R is a strict partial order iff R is the positive walk
relation of a DAG.
Strict partial orders come up in many situations which on the face of it have
nothing to do with digraphs. For example, the less-than order < on numbers is a
strict partial order:
“mcs” — 2017/6/5 — 19:42 — page 401 — #409
10.6. Partial Orders
401
if x < y and y < z then x < z, so less-than is transitive, and
NOT.x < x/, so less-than is irreflexive.
The proper containment relation is also a partial order:
if A B and B C then A C , so containment is transitive, and
NOT.A A/, so proper containment is irreflexive.
If there are two vertices that are reachable from each other, then there is a positive length closed walk that starts at one vertex, goes to the other, and then comes
back. So DAGs are digraphs in which no two vertices are mutually reachable. This
corresponds to a relational property called asymmetry.
Definition 10.6.9. A binary relation R on a set A is asymmetric iff
a R b IMPLIES NOT.b R a/
for all a; b 2 A.
So we can also characterize DAGs in terms of asymmetry:
Corollary 10.6.10. A digraph D is a DAG iff D C is asymmetric.
Corollary 10.6.10 and Theorem 10.6.8 combine to give
Corollary 10.6.11. A binary relation R on a set A is a strict partial order iff it is
transitive and asymmetric.9
A strict partial order may be the positive walk relation of different DAGs. This
raises the question of finding a DAG with the smallest number of edges that determines a given strict partial order. For finite strict partial orders, the smallest such
DAG turns out to be unique and easy to find (see Problem 10.30).
10.6.3
Weak Partial Orders
The less-than-or-equal relation is at least as familiar as the less-than strict partial
order, and the ordinary containment relation is even more common than the
proper containment relation. These are examples of weak partial orders, which are
just strict partial orders with the additional condition that every element is related
to itself. To state this precisely, we have to relax the asymmetry property so it
does not apply when a vertex is compared to itself; this relaxed property is called
antisymmetry:
9 Some
texts use this corollary to define strict partial orders.
“mcs” — 2017/6/5 — 19:42 — page 402 — #410
402
Chapter 10
Directed graphs & Partial Orders
Definition 10.6.12. A binary relation R on a set A, is antisymmetric iff, for all
a ¤ b 2 A,
a R b IMPLIES NOT.b R a/
Now we can give an axiomatic definition of weak partial orders that parallels the
definition of strict partial orders.
Definition 10.6.13. A binary relation on a set is a weak partial order iff it is transitive, reflexive, and antisymmetric.
The following lemma gives another characterization of weak partial orders that
follows directly from this definition.
Lemma 10.6.14. A relation R on a set A is a weak partial order iff there is a strict
partial order S on A such that
aRb
iff
.a S b OR a D b/;
for all a; b 2 A.
Since a length zero walk goes from a vertex to itself, this lemma combined with
Theorem 10.6.8 yields:
Corollary 10.6.15. A relation is a weak partial order iff it is the walk relation of a
DAG.
For weak partial orders in general, we often write an ordering-style symbol like
or v instead of a letter symbol like R.10 Likewise, we generally use or @ to
indicate a strict partial order.
Two more examples of partial orders are worth mentioning:
Example 10.6.16. Let A be some family of sets and define a R b iff a b. Then
R is a strict partial order.
Example 10.6.17. The divisibility relation is a weak partial order on the nonnegative integers.
For practice with the definitions, you can check that two more examples are
vacuously partial orders on a set D: the identity relation IdD is a weak partial
order, and the empty relation—the relation with no arrows—is a strict partial order.
Note that some authors define “partial orders” to be what we call weak partial
orders. However, we’ll use the phrase “partial order” to mean a relation that may
be either a weak or strict partial order.
10 General relations are usually denoted by a letter like R instead of a cryptic squiggly symbol, so
is kind of like the musical performer/composer Prince, who redefined the spelling of his name to
be his own squiggly symbol. A few years ago he gave up and went back to the spelling “Prince.”
“mcs” — 2017/6/5 — 19:42 — page 403 — #411
10.7. Representing Partial Orders by Set Containment
10.7
403
Representing Partial Orders by Set Containment
Axioms can be a great way to abstract and reason about important properties of
objects, but it helps to have a clear picture of the things that satisfy the axioms.
DAGs provide one way to picture partial orders, but it also can help to picture them
in terms of other familiar mathematical objects. In this section, we’ll show that
every partial order can be pictured as a collection of sets related by containment.
That is, every partial order has the “same shape” as such a collection. The technical
word for “same shape” is “isomorphic.”
Definition 10.7.1. A binary relation R on a set A is isomorphic to a relation S on
a set B iff there is a relation-preserving bijection from A to B; that is, there is a
bijection f W A ! B such that for all a; a0 2 A,
a R a0
iff
f .a/ S f .a0 /:
To picture a partial order on a set A as a collection of sets, we simply represent
each element A by the set of elements that are to that element, that is,
! fb 2 A j b ag:
a
For example, if is the divisibility relation on the set of integers f1; 3; 4; 6; 8; 12g,
then we represent each of these integers by the set of integers in A that divides it.
So
1
! f1g
3
! f1; 3g
4
! f1; 4g
6
! f1; 3; 6g
8
! f1; 4; 8g
12
! f1; 3; 4; 6; 12g
So, the fact that 3 j 12 corresponds to the fact that f1; 3g f1; 3; 4; 6; 12g.
In this way we have completely captured the weak partial order by the subset
relation on the corresponding sets. Formally, we have
Lemma 10.7.2. Let be a weak partial order on a set A. Then is isomorphic
to the subset relation on the collection of inverse images under the relation of
elements a 2 A.
“mcs” — 2017/6/5 — 19:42 — page 404 — #412
404
Chapter 10
Directed graphs & Partial Orders
We leave the proof to Problem 10.36. Essentially the same construction shows
that strict partial orders can be represented by sets under the proper subset relation,
(Problem 10.37). To summarize:
Theorem 10.7.3. Every weak partial order is isomorphic to the subset relation
on a collection of sets.
Every strict partial order is isomorphic to the proper subset relation on a
collection of sets.
10.8
Linear Orders
The familiar order relations on numbers have an important additional property:
given two different numbers, one will be bigger than the other. Partial orders with
this property are said to be linear orders. You can think of a linear order as one
where all the elements are lined up so that everyone knows exactly who is ahead
and who is behind them in the line.11
Definition 10.8.1. Let R be a binary relation on a set A and let a; b be elements
of A. Then a and b are comparable with respect to R iff Œa R b OR b R a.
A partial order for which every two different elements are comparable is called a
linear order.
So < and are linear orders on R. On the other hand, the subset relation is
not linear, since, for example, any two different finite sets of the same size will be
incomparable under . The prerequisite relation on Course 6 required subjects is
also not linear because, for example, neither 8.01 nor 6.042 is a prerequisite of the
other.
10.9
Product Orders
Taking the product of two relations is a useful way to construct new relations from
old ones.
11 Linear orders are often called “total” orders, but this terminology conflicts with the definition of
“total relation,” and it regularly confuses students.
Being a linear order is a much stronger condition than being a partial order that is a total relation.
For example, any weak partial order is a total relation but generally won’t be linear.
“mcs” — 2017/6/5 — 19:42 — page 405 — #413
10.10. Equivalence Relations
405
Definition 10.9.1. The product R1 R2 of relations R1 and R2 is defined to be
the relation with
domain.R1 R2 / WWD domain.R1 / domain.R2 /;
codomain.R1 R2 / WWD codomain.R1 / codomain.R2 /;
.a1 ; a2 / .R1 R2 / .b1 ; b2 /
iff
Œa1 R1 b1 and a2 R2 b2 :
It follows directly from the definitions that products preserve the properties of
transitivity, reflexivity, irreflexivity, and antisymmetry (see Problem 10.51). If R1
and R2 both have one of these properties, then so does R1 R2 . This implies that
if R1 and R2 are both partial orders, then so is R1 R2 .
Example 10.9.2. Define a relation Y on age-height pairs of being younger and
shorter. This is the relation on the set of pairs .y; h/ where y is a nonnegative
integer 2400 that we interpret as an age in months, and h is a nonnegative integer
120 describing height in inches. We define Y by the rule
.y1 ; h1 / Y .y2 ; h2 / iff y1 y2 AND h1 h2 :
That is, Y is the product of the -relation on ages and the -relation on heights.
Since both ages and heights are ordered numerically, the age-height relation Y is
a partial order. Now suppose we have a class of 101 students. Then we can apply
Dilworth’s lemma 10.5.12 to conclude that there is a chain of 11 students—that is,
11 students who get taller as they get older–or an antichain of 11 students—that is,
11 students who get taller as they get younger, which makes for an amusing in-class
demo.
On the other hand, the property of being a linear order is not preserved. For
example, the age-height relation Y is the product of two linear orders, but it is not
linear: the age 240 months, height 68 inches pair, (240,68), and the pair (228,72)
are incomparable under Y .
10.10
Equivalence Relations
Definition 10.10.1. A relation R on a set A is symmetric when
8x; y 2 A: x R y IMPLIES y R x:
A relation is an equivalence relation iff it is reflexive, symmetric, and transitive.
Congruence modulo n is an important example of an equivalence relation:
“mcs” — 2017/6/5 — 19:42 — page 406 — #414
406
Chapter 10
Directed graphs & Partial Orders
It is reflexive because x x .mod n/.
It is symmetric because x y .mod n/ implies y x .mod n/.
It is transitive because x y .mod n/ and y z .mod n/ imply that x z
.mod n/.
There is an even more well-known example of an equivalence relation: equality
itself.
If we think of a total function f as attaching a label to each element in its domain,
then “having the same label” defines an equivalence relation. More precisely, if f
is a total function, we can define a basic equivalence relation f on the domain of
f as follows:
Definition 10.10.2. If f is a total function, the relation f on domain.f / is defined by the rule:
a f b IFF f .a/ D f .b/:
The relation f inherits the properties of reflexivity, symmetry and transitivity
from the corresponding properties of the equality relation. So f is indeed an
equivalence relation. This observation gives another way to see that congruence
modulo n is an equivalence relation: the Remainder Lemma 9.6.1 implies that
congruence modulo n is the same as r where r.a/ is the remainder of a divided
by n.
In fact, a relation is an equivalence relation iff it equals f for some total function f (see Problem 10.57). So every equivalence relation can actually be understood as being the relation of “having the same label” according to some labelling
scheme.
10.10.1
Equivalence Classes
Equivalence relations are closely related to partitions because the images of elements under an equivalence relation are the blocks of a partition.
Definition 10.10.3. Given an equivalence relation R W A ! A, the equivalence
class ŒaR of an element a 2 A is the set of all elements of A related to a by R.
Namely,
ŒaR WWD fx 2 A j a R xg:
In other words, ŒaR is the image R.a/.
For example, suppose that A D Z and a R b means that a b .mod 5/. Then
Œ7R D f: : : ; 3; 2; 7; 12; 22; : : :g:
“mcs” — 2017/6/5 — 19:42 — page 407 — #415
10.11. Summary of Relational Properties
407
Notice that 7, 12, 17, etc., all have the same equivalence class; that is, Œ7R D
Œ12R D Œ17R D .
There is an exact correspondence between equivalence relations on A and partitions of A. Namely, given any partition of a set, being in the same block is obviously
an equivalence relation. On the other hand we have:
Theorem 10.10.4. The equivalence classes of an equivalence relation on a set A
are the blocks of a partition of A.
We’ll leave the proof of Theorem 10.10.4 as a basic exercise in axiomatic reasoning (see Problem 10.56), but let’s look at an example. The congruent-mod-5
relation partitions the integers into five equivalence classes:
f: : : ; 5; 0; 5; 10; 15; 20; : : :g
f: : : ; 4; 1; 6; 11; 16; 21; : : :g
f: : : ; 3; 2; 7; 12; 17; 22; : : :g
f: : : ; 2; 3; 8; 13; 18; 23; : : :g
f: : : ; 1; 4; 9; 14; 19; 24; : : :g
In these terms, x y .mod 5/ is equivalent to the assertion that x and y are both
in the same block of this partition. For example, 6 16 .mod 5/, because they’re
both in the second block, but 2 6 9 .mod 5/ because 2 is in the third block while
9 is in the last block.
In social terms, if “likes” were an equivalence relation, then everyone would be
partitioned into cliques of friends who all like each other and no one else.
10.11
Summary of Relational Properties
A relation R W A ! A is the same as a digraph with vertices A.
Reflexivity R is reflexive when
8x 2 A: x R x:
Every vertex in R has a self-loop.
Irreflexivity R is irreflexive when
NOT Œ9x
There are no self-loops in R.
2 A: x R x:
“mcs” — 2017/6/5 — 19:42 — page 408 — #416
408
Chapter 10
Directed graphs & Partial Orders
Symmetry R is symmetric when
8x; y 2 A: x R y IMPLIES y R x:
If there is an edge from x to y in R, then there is an edge back from y to x
as well.
Asymmetry R is asymmetric when
8x; y 2 A: x R y IMPLIES NOT.y R x/:
There is at most one directed edge between any two vertices in R, and there
are no self-loops.
Antisymmetry R is antisymmetric when
8x ¤ y 2 A: x R y IMPLIES NOT.y R x/:
Equivalently,
8x; y 2 A: .x R y AND y R x/ IMPLIES x D y:
There is at most one directed edge between any two distinct vertices, but
there may be self-loops.
Transitivity R is transitive when
8x; y; z 2 A: .x R y AND y R z/ IMPLIES x R z:
If there is a positive length path from u to v, then there is an edge from u
to v.
Linear R is linear when
8x ¤ y 2 A: .x R y OR y R x/
Given any two vertices in R, there is an edge in one direction or the other
between them.
Strict Partial Order R is a strict partial order iff R is transitive and irreflexive iff
R is transitive and asymmetric iff it is the positive length walk relation of a
DAG.
Weak Partial Order R is a weak partial order iff R is transitive and anti-symmetric
and reflexive iff R is the walk relation of a DAG.
Equivalence Relation R is an equivalence relation iff R is reflexive, symmetric
and transitive iff R equals the in-the-same-block-relation for some partition
of domain.R/.
“mcs” — 2017/6/5 — 19:42 — page 409 — #417
10.12. References
10.12
409
References
[9], [15], [24], [27], [29]
Problems for Section 10.1
Practice Problems
Problem 10.1.
Let S be a nonempty set of size n 2 ZC , and let f W S ! S be total function. Let
Df be the digraph with vertices S whose edges are fhs ! f .s/i j s 2 S g.
(a) What are the possible values of the out-degrees of vertices of Df ?
(b) What are the possible values of the in-degrees of the vertices?
(c) Suppose f is a surjection. Now what are the possible values of the in-degrees
of the vertices?
Exam Problems
Problem 10.2.
The proof of the Handshaking Lemma 10.1.2 invoked the “obvious” fact that in any
finite digraph, the sum of the in-degrees of the vertices equals the number of arrows
in the graph. That is,
Claim. For any finite digraph G
X
indeg.v/ D j graph.G/j;
(10.10)
v2V .G/
But this Claim might not be obvious to everyone. So prove it by induction on the
number j graph.G/j of arrows.
Problems for Section 10.2
Practice Problems
Problem 10.3.
Lemma 10.2.5 states that dist .u; v/ dist .u; x/ C dist .x; v/. It also states that
“mcs” — 2017/6/5 — 19:42 — page 410 — #418
410
Chapter 10
Directed graphs & Partial Orders
equality holds iff x is on a shortest path from u to v.
(a) Prove the “iff” statement from left to right.
(b) Prove the “iff” from right to left.
Class Problems
Problem 10.4. (a) Give an example of a digraph that has a closed walk including
two vertices but has no cycle including those vertices.
(b) Prove Lemma 10.2.6:
Lemma. The shortest positive length closed walk through a vertex is a cycle.
Problem 10.5.
A 3-bit string is a string made up of 3 characters, each a 0 or a 1. Suppose you’d
like to write out, in one string, all eight of the 3-bit strings in any convenient order.
For example, if you wrote out the 3-bit strings in the usual order starting with 000
001 010. . . , you could concatenate them together to get a length 3 8 D 24 string
that started 000001010. . . .
But you can get a shorter string containing all eight 3-bit strings by starting with
00010. . . . Now 000 is present as bits 1 through 3, and 001 is present as bits 2
through 4, and 010 is present as bits 3 through 5, . . . .
(a) Say a string is 3-good if it contains every 3-bit string as 3 consecutive bits
somewhere in it. Find a 3-good string of length 10, and explain why this is the
minimum length for any string that is 3-good.
(b) Explain how any walk that includes every edge in the graph shown in Figure 10.10 determines a string that is 3-good. Find the walk in this graph that determines your 3-good string from part (a).
(c) Explain why a walk in the graph of Figure 10.10 that includes every every
edge exactly once provides a minimum-length 3-good string.12
(d) Generalize the 2-bit graph to a k-bit digraph Bk for k 2, where V .Bk / WWD
f0; 1gk , and any walk through Bk that contains every edge exactly once determines
a minimum length .k C 1/-good bit-string.13
12 The
3-good strings explained here generalize to n-good strings for n 3. They were studied by
the great Dutch mathematician/logician Nicolaas de Bruijn, and are known as de Bruijn sequences.
de Bruijn died in February, 2012 at the age of 94.
13 Problem 10.7 explains why such “Eulerian” paths exist.
“mcs” — 2017/6/5 — 19:42 — page 411 — #419
10.12. References
411
+1
10
+0
11
+1
+0
+1
+0
00
01
+0
+1
Figure 10.10 The 2-bit graph.
What is this minimum length?
Define the transitions of Bk . Verify that the in-degree of each vertex is the same as
its out-degree and that there is a positive length path from any vertex to any other
vertex (including itself) of length at most k.
Homework Problems
Problem 10.6. (a) Give an example of a digraph in which a vertex v is on a positive
even-length closed walk, but no vertex is on an even-length cycle.
(b) Give an example of a digraph in which a vertex v is on an odd-length closed
walk but not on an odd-length cycle.
(c) Prove that every odd-length closed walk contains a vertex that is on an oddlength cycle.
Problem 10.7.
An Euler tour14 of a graph is a closed walk that includes every edge exactly once.
14 In
some other texts, this is called an Euler circuit.
“mcs” — 2017/6/5 — 19:42 — page 412 — #420
412
Chapter 10
Directed graphs & Partial Orders
Such walks are named after the famous 17th century mathematician Leonhard Euler. (Same Euler as for the constant e 2:718 and the totient function —he did a
lot of stuff.)
So how do you tell in general whether a graph has an Euler tour? At first glance
this may seem like a daunting problem (the similar sounding problem of finding
a cycle that touches every vertex exactly once is one of those million dollar NPcomplete problems known as the Hamiltonian Cycle Problem)—but it turns out to
be easy.
(a) Show that if a graph has an Euler tour, then the in-degree of each vertex equals
its out-degree.
A digraph is weakly connected if there is a “path” between any two vertices that
may follow edges backwards or forwards.15 In the remaining parts, we’ll work out
the converse. Suppose a graph is weakly connected, and the in-degree of every
vertex equals its out-degree. We will show that the graph has an Euler tour.
A trail is a walk in which each edge occurs at most once.
(b) Suppose that a trail in a weakly connected graph does not include every edge.
Explain why there must be an edge not on the trail that starts or ends at a vertex on
the trail.
In the remaining parts, assume the graph is weakly connected, and the in-degree
of every vertex equals its out-degree. Let w be the longest trail in the graph.
(c) Show that if w is closed, then it must be an Euler tour.
Hint: part (b)
(d) Explain why all the edges starting at the end of w must be on w.
(e) Show that if w was not closed, then the in-degree of the end would be bigger
than its out-degree.
Hint: part (d)
(f) Conclude that if the in-degree of every vertex equals its out-degree in a finite,
weakly connected digraph, then the digraph has an Euler tour.
15 More precisely, a graph G is weakly connected iff there is a path from any vertex to any other
vertex in the graph H with
V .H / D V .G/; and
E.H / D E.G/ [ fhv ! ui j hu ! vi 2 E.G/g:
In other words H D G [ G
1.
“mcs” — 2017/6/5 — 19:42 — page 413 — #421
10.12. References
413
Problems for Section 10.3
Homework Problems
Problem 10.8.
The weight of a walk in a weighted graph is the sum of the weights of the successive
edges in the walk. The minimum weight matrix for length k walks in an n-vertex
graph G is the n n matrix W such that for u; v 2 V .G/,
(
w if w is the minimum weight among length k walks from u to v;
Wuv WWD
1 if there is no length k walk from u to v:
The min+ product of two n n matrices W and M with entries in R [ f1g is the
n n matrix W M whose ij entry is
min+
.W V /ij WWD minfWi k C Vkj j 1 k ng :
min+
Prove the following theorem.
Theorem. If W is the minimum weight matrix for length k walks in a weighted
graph G, and V is the minimum weight matrix for length m walks, then W V is
min+
the minimum weight matrix for length k C m walks.
Problems for Section 10.4
Practice Problems
Problem 10.9.
Let
A WWD f1; 2; 3g
B WWD f4; 5; 6g
R WWD f.1; 4/; .1; 5/; .2; 5/; .3; 6/g
S WWD f.4; 5/; .4; 6/; .5; 4/g:
Note that R is a relation from A to B and S is a relation from B to B.
List the pairs in each of the relations below.
“mcs” — 2017/6/5 — 19:42 — page 414 — #422
414
Chapter 10
Directed graphs & Partial Orders
(a) S ı R.
(b) S ı S .
(c) S
1
ı R.
Problem 10.10.
In a round-robin tournament, every two distinct players play against each other
just once. For a round-robin tournament with no tied games, a record of who beat
whom can be described with a tournament digraph, where the vertices correspond
to players and there is an edge hx ! yi iff x beat y in their game.
A ranking is a path that includes all the players. So in a ranking, each player won
the game against the next lowest ranked player, but may very well have lost their
games against much lower ranked players—whoever does the ranking may have a
lot of room to play favorites.
(a) Give an example of a tournament digraph with more than one ranking.
(b) Prove that if a tournament digraph is a DAG, then it has at most one ranking.
(c) Prove that every finite tournament digraph has a ranking.
Optional
(d) Prove that the greater-than relation > on the rational numbers Q is a DAG and
a tournament graph that has no ranking.
Homework Problems
Problem 10.11.
Let R be a binary relation on a set A. Regarding R as a digraph, let W .n/ denote
the length-n walk relation in the digraph R, that is,
a W .n/ b WWD there is a length n walk from a to b in R:
(a) Prove that
W .n/ ı W .m/ D W .mCn/
(10.11)
for all m; n 2 N, where ı denotes relational composition.
(b) Let Rn be the composition of R with itself n times for n 0. So R0 WWD IdA ,
and RnC1 WWD R ı Rn .
“mcs” — 2017/6/5 — 19:42 — page 415 — #423
10.12. References
415
Conclude that
Rn D W .n/
(10.12)
for all n 2 N.
(c) Conclude that
RC D
jAj
[
Ri
i D1
where
RC
is the positive length walk relation determined by R on the set A.
Problem 10.12.
We can represent a relation S between two sets A D fa1 ; : : : ; an g and B D
fb1 ; : : : ; bm g as an n m matrix MS of zeroes and ones, with the elements of
MS defined by the rule
MS .i; j / D 1
IFF
ai S bj :
If we represent relations as matrices this way, then we can compute the composition of two relations R and S by a “boolean” matrix multiplication ˝ of their
matrices. Boolean matrix multiplication is the same as matrix multiplication except
that addition is replaced by OR, multiplication is replaced by AND, and 0 and 1 are
used as the Boolean values False and True. Namely, suppose R W B ! C is a binary relation with C D fc1 ; : : : ; cp g. So MR is an m p matrix. Then MS ˝ MR
is an n p matrix defined by the rule:
ŒMS ˝ MR .i; j / WWD ORm
kD1 ŒMS .i; k/ AND MR .k; j /:
(10.13)
Prove that the matrix representation MRıS of R ı S equals MS ˝ MR (note the
reversal of R and S).
Problem 10.13.
Chickens are rather aggressive birds that tend to establish dominance over other
chickens by pecking them—hence the term “pecking order.” So for any two chickens in a farmyard, either the first pecks the second, or the second pecks the first, but
not both. We say that chicken u virtually pecks chicken v if either:
Chicken u pecks chicken v, or
Chicken u pecks some other chicken w who in turn pecks chicken v.
“mcs” — 2017/6/5 — 19:42 — page 416 — #424
416
Chapter 10
Directed graphs & Partial Orders
A chicken that virtually pecks every other chicken is called a king chicken.
We can model this situation with a chicken digraph whose vertices are chickens,
with an edge from chicken u to chicken v precisely when u pecks v. In the graph
in Figure 10.11, three of the four chickens are kings. Chicken c is not a king in
this example since it does not peck chicken b and it does not peck any chicken that
pecks chicken b. Chicken a is a king since it pecks chicken d , who in turn pecks
chickens b and c.
In general, a tournament digraph is a digraph with exactly one edge between
each pair of distinct vertices.
king
king
a
b
d
c
king
not a king
Figure 10.11 A 4-chicken tournament in which chickens a, b and d are kings.
.
(a) Define a 10-chicken tournament graph with a king chicken that has outdegree
1.
(b) Describe a 5-chicken tournament graph in which every player is a king.
(c) Prove
Theorem (King Chicken Theorem). Any chicken with maximum out-degree in a
tournament is a king.
The King Chicken Theorem means that if the player with the most victories is
defeated by another player x, then at least he/she defeats some third player that
defeats x. In this sense, the player with the most victories has some sort of bragging
rights over every other player. Unfortunately, as Figure 10.11 illustrates, there can
be many other players with such bragging rights, even some with fewer victories.
“mcs” — 2017/6/5 — 19:42 — page 417 — #425
10.12. References
417
Problems for Section 10.5
Practice Problems
Problem 10.14.
What is the size of the longest chain that is guaranteed to exist in any partially
ordered set of n elements? What about the largest antichain?
Problem 10.15.
Let fA; :::; H g be a set of tasks that we must complete. The following DAG describes which tasks must be done before others, where there is an arrow from S to
T iff S must be done before T .
(a) Write the longest chain.
(b) Write the longest antichain.
(c) If we allow parallel scheduling, and each task takes 1 minute to complete,
what is the minimum amount of time needed to complete all tasks?
Problem 10.16.
Describe a sequence consisting of the integers from 1 to 10,000 in some order so
that there is no increasing or decreasing subsequence of size 101.
“mcs” — 2017/6/5 — 19:42 — page 418 — #426
418
Chapter 10
Directed graphs & Partial Orders
Problem 10.17.
Suppose the vertices of a DAG represent tasks taking unit time to complete, and
the edges indicate prerequisites among the tasks. Assume there is no bound on how
many tasks may be performed in parallel.
What is the smallest number of vertices (tasks) possible in a DAG for which there
is more than one minimum time schedule? Carefully justify your answer.
Problem 10.18.
The following DAG describes the prerequisites among tasks f1; : : : ; 9g.
9
8
3
5
2
7
4
6
1
(a) If each task takes unit time to complete, what is the minimum parallel time to
complete all the tasks? Briefly explain.
(b) What is the minimum parallel time if no more than two tasks can be completed
in parallel? Briefly explain.
Problem 10.19.
The following DAG describes the prerequisites among tasks f1; : : : ; 9g.
(a) If each task takes unit time to complete, what is the minimum time to complete
all the tasks? Briefly explain.
(b) What is the minimum time if no more than two tasks can be completed in
parallel? Briefly explain.
“mcs” — 2017/6/5 — 19:42 — page 419 — #427
10.12. References
419
8
3
9
5
2
7
4
6
1
Class Problems
Problem 10.20.
The table below lists some prerequisite information for some subjects in the MIT
Computer Science program (in 2006). This defines an indirect prerequisite relation
that is a DAG with these subjects as vertices.
18:01 ! 6:042
18:01 ! 18:02
18:01 ! 18:03
6:046 ! 6:840
8:01 ! 8:02
6:001 ! 6:034
6:042 ! 6:046
18:03; 8:02 ! 6:002
6:001; 6:002 ! 6:003
6:001; 6:002 ! 6:004
6:004 ! 6:033
6:033 ! 6:857
(a) Explain why exactly six terms are required to finish all these subjects, if you
can take as many subjects as you want per term. Using a greedy subject selection
strategy, you should take as many subjects as possible each term. Exhibit your
complete class schedule each term using a greedy strategy.
(b) In the second term of the greedy schedule, you took five subjects including
18.03. Identify a set of five subjects not including 18.03 such that it would be
possible to take them in any one term (using some nongreedy schedule). Can you
figure out how many such sets there are?
“mcs” — 2017/6/5 — 19:42 — page 420 — #428
420
Chapter 10
Directed graphs & Partial Orders
(c) Exhibit a schedule for taking all the courses—but only one per term.
(d) Suppose that you want to take all of the subjects, but can handle only two per
term. Exactly how many terms are required to graduate? Explain why.
(e) What if you could take three subjects per term?
Problem 10.21.
A pair of Math for Computer Science Teaching Assistants, Lisa and Annie, have
decided to devote some of their spare time this term to establishing dominion over
the entire galaxy. Recognizing this as an ambitious project, they worked out the
following table of tasks on the back of Annie’s copy of the lecture notes.
1. Devise a logo and cool imperial theme music - 8 days.
2. Build a fleet of Hyperwarp Stardestroyers out of eating paraphernalia swiped
from Lobdell - 18 days.
3. Seize control of the United Nations - 9 days, after task #1.
4. Get shots for Lisa’s cat, Tailspin - 11 days, after task #1.
5. Open a Starbucks chain for the army to get their caffeine - 10 days, after
task #3.
6. Train an army of elite interstellar warriors by dragging people to see The
Phantom Menace dozens of times - 4 days, after tasks #3, #4, and #5.
7. Launch the fleet of Stardestroyers, crush all sentient alien species, and establish a Galactic Empire - 6 days, after tasks #2 and #6.
8. Defeat Microsoft - 8 days, after tasks #2 and #6.
We picture this information in Figure 10.12 below by drawing a point for each
task, and labelling it with the name and weight of the task. An edge between
two points indicates that the task for the higher point must be completed before
beginning the task for the lower one.
(a) Give some valid order in which the tasks might be completed.
Lisa and Annie want to complete all these tasks in the shortest possible time.
However, they have agreed on some constraining work rules.
Only one person can be assigned to a particular task; they cannot work together on a single task.
“mcs” — 2017/6/5 — 19:42 — page 421 — #429
10.12. References
421
devise logo
u8
A
A uget
11
B
B
B
B
B
B
shots
B
launch fleet
Figure 10.12
E
E
E
E
E
E
B
Q
4
E
QQ
B E
army B uP
QPP
E
Q PP
PP Q
E
PP
Q
E
PP
Q
Q PP
EE defeat
Pu
Q
u
6
Q
Q
train
E
E
E
E
E
E
E
A
A
A
seize control u9
open chain u
10 QQ
build fleet
u 18
Microsoft
8
Graph representing the task precedence constraints.
Once a person is assigned to a task, that person must work exclusively on
the assignment until it is completed. So, for example, Lisa cannot work on
building a fleet for a few days, run to get shots for Tailspin, and then return
to building the fleet.
(b) Lisa and Annie want to know how long conquering the galaxy will take. Annie
suggests dividing the total number of days of work by the number of workers, which
is two. What lower bound on the time to conquer the galaxy does this give, and why
might the actual time required be greater?
(c) Lisa proposes a different method for determining the duration of their project.
She suggests looking at the duration of the critical path, the most time-consuming
sequence of tasks such that each depends on the one before. What lower bound
does this give, and why might it also be too low?
(d) What is the minimum number of days that Lisa and Annie need to conquer the
galaxy? No proof is required.
Problem 10.22.
“mcs” — 2017/6/5 — 19:42 — page 422 — #430
422
Chapter 10
Directed graphs & Partial Orders
Answer the following questions about the powerset pow.f1; 2; 3; 4g/ partially ordered by the strict subset relation .
(a) Give an example of a maximum length chain.
(b) Give an example of an antchain of size 6.
(c) Describe an example of a topological sort of pow.f1; 2; 3; 4g/.
(d) Suppose the partial order describes scheduling constraints on 16 tasks. That
is, if
A B f1; 2; 3; 4g;
then A has to be completed before B starts.16 What is the minimum number of
processors needed to complete all the tasks in minimum parallel time?
Prove it.
(e) What is the length of a minimum time 3-processor schedule?
Prove it.
Homework Problems
Problem 10.23.
The following operations can be applied to any digraph, G:
1. Delete an edge that is in a cycle.
2. Delete edge hu ! vi if there is a path from vertex u to vertex v that does not
include hu ! vi.
3. Add edge hu ! vi if there is no path in either direction between vertex u and
vertex v.
The procedure of repeating these operations until none of them are applicable can
be modeled as a state machine. The start state is G, and the states are all possible
digraphs with the same vertices as G.
(a) Let G be the graph with vertices f1; 2; 3; 4g and edges
fh1 ! 2i ; h2 ! 3i ; h3 ! 4i ; h3 ! 2i ; h1 ! 4ig
What are the possible final states reachable from G?
A line graph is a graph whose edges are all on one path. All the final graphs in
part (a) are line graphs.
16 As
usual, we assume each task requires one time unit to complete.
“mcs” — 2017/6/5 — 19:42 — page 423 — #431
10.12. References
423
(b) Prove that if the procedure terminates with a digraph H then H is a line graph
with the same vertices as G.
Hint: Show that if H is not a line graph, then some operation must be applicable.
(c) Prove that being a DAG is a preserved invariant of the procedure.
(d) Prove that if G is a DAG and the procedure terminates, then the walk relation
of the final line graph is a topological sort of G.
Hint: Verify that the predicate
P .u; v/ WWD there is a directed path from u to v
is a preserved invariant of the procedure, for any two vertices u; v of a DAG.
(e) Prove that if G is finite, then the procedure terminates.
Hint: Let s be the number of cycles, e be the number of edges, and p be the number
of pairs of vertices with a directed path (in either direction) between them. Note
that p n2 where n is the number of vertices of G. Find coefficients a; b; c such
that as C bp C e C c is nonnegative integer valued and decreases at each transition.
Problem 10.24.
Let be a strict partial order on a set A and let
Ak WWD fa j depth .a/ D kg
where k 2 N.
(a) Prove that A0 ; A1 ; : : : is a parallel schedule for according to Definition 10.5.7.
(b) Prove that Ak is an antichain.
Problem 10.25.
We want to schedule n tasks with prerequisite constraints among the tasks defined
by a DAG.
(a) Explain why any schedule that requires only p processors must take time at
least dn=pe.
(b) Let Dn;t be the DAG with n elements that consists of a chain of t 1 elements,
with the bottom element in the chain being a prerequisite of all the remaining elements as in the following figure:
“mcs” — 2017/6/5 — 19:42 — page 424 — #432
Chapter 10
Directed graphs & Partial Orders
t-1
...
424
...
n - (t - 1)
What is the minimum time schedule for Dn;t ? Explain why it is unique. How many
processors does it require?
(c) Write a simple formula M.n; t; p/ for the minimum time of a p-processor
schedule to complete Dn;t .
(d) Show that every partial order with n vertices and maximum chain size t has a
p-processor schedule that runs in time M.n; t; p/.
Hint: Use induction on t.
Problems for Section 10.6
Practice Problems
Problem 10.26.
In this DAG (Figure 10.13) for the divisibility relation on f1; : : : ; 12g, there is
an upward path from a to b iff a j b. If 24 was added as a vertex, what is the
minimum number of edges that must be added to the DAG to represent divisibility
on set 1; : : : ; 12; 24? What are those edges?
Problem 10.27. (a) Prove that every strict partial order is a DAG.
“mcs” — 2017/6/5 — 19:42 — page 425 — #433
10.12. References
425
8
12
4
9
6
11
3
10
2
5
7
1
Figure 10.13
(b) What is the smallest possible size of a DAG that is not transitive? Prove it.
(c) Prove that the positive walk relation of a DAG a strict partial order.
Class Problems
Problem 10.28. (a) What are the maximal and minimal elements, if any, of the
power set pow.f1; : : : ; ng/, where n is a positive integer, under the empty relation?
(b) What are the maximal and minimal elements, if any, of the set N of all nonnegative integers under divisibility? Is there a minimum or maximum element?
(c) What are the minimal and maximal elements, if any, of the set of integers
greater than 1 under divisibility?
(d) Describe a partially ordered set that has no minimal or maximal elements.
(e) Describe a partially ordered set that has a unique minimal element, but no
minimum element. Hint: It will have to be infinite.
Problem 10.29.
The proper subset relation defines a strict partial order on the subsets of Œ1::6,
that is, on pow.Œ1::6/.
(a) What is the size of a maximal chain in this partial order? Describe one.
(b) Describe a largest antichain in this partial order.
(c) What are the maximal and minimal elements? Are they maximum and minimum?
(d) Answer the previous part for the partial order on the set pow Œ1::6
;.
“mcs” — 2017/6/5 — 19:42 — page 426 — #434
426
Chapter 10
Directed graphs & Partial Orders
Problem 10.30.
If a and b are distinct nodes of a digraph, then a is said to cover b if there is an
edge from a to b and every path from a to b includes this edge. If a covers b, the
edge from a to b is called a covering edge.
(a) What are the covering edges in the DAG in Figure 10.14?
(b) Let covering .D/ be the subgraph of D consisting of only the covering edges.
Suppose D is a finite DAG. Explain why covering .D/ has the same positive walk
relation as D.
Hint: Consider longest paths between a pair of vertices.
(c) Show that if two DAG’s have the same positive walk relation, then they have
the same set of covering edges.
(d) Conclude that covering .D/ is the unique DAG with the smallest number of
edges among all digraphs with the same positive walk relation as D.
The following examples show that the above results don’t work in general for
digraphs with cycles.
(e) Describe two graphs with vertices f1; 2g which have the same set of covering
edges, but not the same positive walk relation (Hint: Self-loops.)
(f)
(i) The complete digraph without self-loops on vertices 1; 2; 3 has directed
edges in each direction between every two distinct vertices. What are its
covering edges?
(ii) What are the covering edges of the graph with vertices 1; 2; 3 and edges
h1 ! 2i ; h2 ! 3i ; h3 ! 1i?
(iii) What about their positive walk relations?
Problems for Section 10.6
Exam Problems
Problem 10.31.
Prove that for any nonempty set D, there is a unique binary relation on D that is
both asymmetric and symmetric.
“mcs” — 2017/6/5 — 19:42 — page 427 — #435
10.12. References
427
2
4
6
1
3
5
Figure 10.14 DAG with edges not needed in paths
Problem 10.32.
Let D be a set of size n > 0. Explain why there are exactly 2n binary relations on
D that are both symmetric and antisymmetric.
Homework Problems
Problem 10.33.
Prove that if R is a transitive binary relation on a set A then R D RC .
Class Problems
Problem 10.34.
Let R be a binary relation on a set D. Each of the following equalities and containments expresses the fact that R has one of the basic relational properties: reflexive,
irreflexive, symmetric, asymmetric, antisymmetric, transitive. Identify which property is expressed by each of these formulas and explain your reasoning.
(a) R \ IdD D ;
(b) R R
1
(c) R D R
1
(d) IdD R
(e) R ı R R
“mcs” — 2017/6/5 — 19:42 — page 428 — #436
428
Chapter 10
Directed graphs & Partial Orders
(f) R \ R
1
D;
(g) R \ R
1
IdD
Problems for Section 10.7
Class Problems
Problem 10.35.
Direct Prerequisites
18.01
18.01
18.01
8.01
8.01
6.042
18.02, 18.03, 8.02, 6.01
6.01, 6.042
6.01
6.02
Subject
6.042
18.02
18.03
8.02
6.01
6.046
6.02
6.006
6.034
6.004
(a) For the above table of MIT subject prerequisites, draw a diagram showing the
subject numbers with a line going down to every subject from each of its (direct)
prerequisites.
(b) Give an example of a collection of sets partially ordered by the proper subset
relation that is isomorphic to (“same shape as”) the prerequisite relation among
MIT subjects from part (a).
(c) Explain why the empty relation is a strict partial order and describe a collection
of sets partially ordered by the proper subset relation that is isomorphic to the empty
relation on five elements—that is, the relation under which none of the five elements
is related to anything.
(d) Describe a simple collection of sets partially ordered by the proper subset relation that is isomorphic to the ”properly contains” relation on pow f1; 2; 3; 4g.
“mcs” — 2017/6/5 — 19:42 — page 429 — #437
10.12. References
429
Problem 10.36.
This problem asks for a proof of Lemma 10.7.2 showing that every weak partial
order can be represented by (is isomorphic to) a collection of sets partially ordered
under set inclusion (). Namely,
Lemma. Let be a weak partial order on a set A. For any element a 2 A, let
L.a/ WWD fb 2 A j b ag;
L WWD fL.a/ j a 2 Ag:
Then the function L W A ! L is an isomorphism from the relation on A, to the
subset relation on L.
(a) Prove that the function L W A ! L is a bijection.
(b) Complete the proof by showing that
ab
iff
L.a/ L.b/
(10.14)
for all a; b 2 A.
Homework Problems
Problem 10.37.
Every partial order is isomorphic to a collection of sets under the subset relation
(see Section 10.7). In particular, if R is a strict partial order on a set A and a 2 A,
define
L.a/ WWD fag [ fx 2 A j x R ag:
(10.15)
Then
aRb
iff
L.a/ L.b/
(10.16)
holds for all a; b 2 A.
(a) Carefully prove statement (10.16), starting from the definitions of strict partial
order and the strict subset relation .
(b) Prove that if L.a/ D L.b/ then a D b.
(c) Give an example showing that the conclusion of part (b) would not hold if the
definition of L.a/ in equation (10.15) had omitted the expression “fag[.”
“mcs” — 2017/6/5 — 19:42 — page 430 — #438
430
Chapter 10
Directed graphs & Partial Orders
Problems for Section 10.8
Practice Problems
Problem 10.38.
For each of the binary relations below, state whether it is a strict partial order, a
weak partial order, or neither. If it is not a partial order, indicate which of the
axioms for partial order it violates.
(a) The superset relation, on the power set pow f1; 2; 3; 4; 5g.
(b) The relation between any two nonnegative integers a, b given by a b
.mod 8/.
(c) The relation between propositional formulas G, H given by G IMPLIES H is
valid.
(d) The relation ’beats’ on Rock, Paper and Scissor (for those who don’t know the
game “Rock, Paper, Scissors:” Rock beats Scissors, Scissors beats Paper and Paper
beats Rock).
(e) The empty relation on the set of real numbers.
(f) The identity relation on the set of integers.
Problem 10.39. (a) Verify that the divisibility relation on the set of nonnegative
integers is a weak partial order.
(b) What about the divisibility relation on the set of integers?
Problem 10.40.
Prove directly from the definitions (without appealing to DAG properties) that if a
binary relation R on a set A is transitive and irreflexive, then it is asymmetric.
Class Problems
Problem 10.41.
Show that the set of nonnegative integers partially ordered under the divides relation. . .
(a) . . . has a minimum element.
“mcs” — 2017/6/5 — 19:42 — page 431 — #439
10.12. References
431
(b) . . . has a maximum element.
(c) . . . has an infinite chain.
(d) . . . has an infinite antichain.
(e) What are the minimal elements of divisibility on the integers greater than 1?
What are the maximal elements?
Problem 10.42.
How many binary relations are there on the set f0; 1g?
How many are there that are transitive?, . . . asymmetric?, . . . reflexive?, . . . irreflexive?,
. . . strict partial orders?, . . . weak partial orders?
Hint: There are easier ways to find these numbers than listing all the relations
and checking which properties each one has.
Problem 10.43.
Prove that if R is a partial order, then so is R
1.
Problem 10.44. (a) Indicate which of the following relations below are equivalence relations, (Eq), strict partial orders (SPO), weak partial orders (WPO). For
the partial orders, also indicate whether it is linear (Lin).
If a relation is none of the above, indicate whether it is transitive (Tr), symmetric
(Sym), or asymmetric (Asym).
(i) The relation a D b C 1 between integers a, b,
(ii) The superset relation on the power set of the integers.
(iii) The empty relation on the set of rationals.
(iv) The divides relation on the nonegative integers N.
(v) The divides relation on all the integers Z.
(vi) The divides relation on the positive powers of 4.
(vii) The relatively prime relation on the nonnegative integers.
(viii) The relation “has the same prime factors” on the integers.
“mcs” — 2017/6/5 — 19:42 — page 432 — #440
432
Chapter 10
Directed graphs & Partial Orders
(b) A set of functions f; g W D ! R can be partially ordered by the relation,
where
Œf g WWD 8d 2 D: f .d / g.d /:
Let L be the set of functions f W R ! R of the form
f .x/ D ax C b
for constants a; b 2 R.
Describe an infinite chain and an infinite anti-chain in L.
Problem 10.45.
In an n-player round-robin tournament, every pair of distinct players compete in a
single game. Assume that every game has a winner—there are no ties. The results
of such a tournament can then be represented with a tournament digraph where the
vertices correspond to players and there is an edge hx ! yi iff x beat y in their
game.
(a) Briefly explain why a tournament digraph cannot have cycles of length one or
two.
(b) Briefly explain whether the “beats” relation of a tournament graph always/sometimes/never. . .
. . . is asymmetric.
. . . is reflexive.
. . . is irreflexive.
. . . is transitive.
. . . has a unique directed path including all the vertices.
(c) Prove that a tournament graph is a partial order iff there are no cycles of length
three.
Homework Problems
Problem 10.46.
Let R and S be transitive binary relations on the same set A. Which of the following
new relations must also be transitive? For each part, justify your answer with a brief
argument if the new relation is transitive or with a counterexample if it is not.
(a) R 1
(b) R \ S
“mcs” — 2017/6/5 — 19:42 — page 433 — #441
10.12. References
433
(c) R ı R
(d) R ı S
Problem 10.47.
Digraph G has the unique path property when, for any finite nonempty set of vertices of G, there is a unique directed path going through exactly these vertices.
(a) Prove that if G is a linear strict partial order, then G has the unique path
property.
(b) Prove conversely that if G has the unique path property, then the positive path
relation of G is a linear strict partial order.
Exam Problems
Problem 10.48.
Suppose the precedence constraints on a set of 32 unit time tasks was isomorphic
to the powerset, pow.f1; 2; 3; 4; 5g/ under the strict subset relation .
For example, the task corresponding to the set f2; 4g must be completed before the task corresponding to the set f1; 2; 4g because f2; 4g f1; 2; 4g; the task
corresponding to the empty set must be scheduled first because ; S for every
nonempty set S f1; 2; 3; 4; 5g.
(a) What is the minimum parallel time to complete these tasks?
(b) Describe a maximum size antichain in this partial order.
(c) Briefly explain why the minimum number of processors required to complete
these tasks in minimum parallel time is equal to the size of the maximum antichain.
Problem 10.49.
Let R be a weak partial order on a set A. Suppose C is a finite chain.17
(a) Prove that C has a maximum element. Hint: Induction on the size of C .
(b) Conclude that there is a unique sequence of all the elements of C that is strictly
increasing.
Hint: Induction on the size of C , using part (a).
17 A set C is a chain when it is nonempty, and all elements c; d 2 C are comparable. Elements c
and d are comparable iff Œc R d OR d R c.
“mcs” — 2017/6/5 — 19:42 — page 434 — #442
434
Chapter 10
Directed graphs & Partial Orders
Problems for Section 10.9
Practice Problems
Problem 10.50.
Verify that if either of R1 or R2 is irreflexive, then so is R1 R2 .
Class Problems
Problem 10.51.
Let R1 , R2 be binary relations on the same set A. A relational property is preserved
under product, if R1 R2 has the property whenever both R1 and R2 have the
property.
(a) Verify that each of the following properties are preserved under product.
1. reflexivity,
2. antisymmetry,
3. transitivity.
(b) Verify that if R1 and R2 are partial orders and at least one of them is strict,
then R1 R2 is a strict partial order.
Problem 10.52.
A partial order on a set A is well founded when every non-empty subset of A has a
minimal element. For example, the less-than relation on a well ordered set of real
numbers (see 2.4) is a linear order that is well founded.
Prove that if R and S are well founded partial orders, then so is their product
R S.
Homework Problems
Problem 10.53.
Let S be a sequence of n different numbers. A subsequence of S is a sequence that
can be obtained by deleting elements of S .
For example, if S is
.6; 4; 7; 9; 1; 2; 5; 3; 8/;
then 647 and 7253 are both subsequences of S (for readability, we have dropped the
parentheses and commas in sequences, so 647 abbreviates .6; 4; 7/, for example).
“mcs” — 2017/6/5 — 19:42 — page 435 — #443
10.12. References
435
An increasing subsequence of S is a subsequence of whose successive elements
get larger. For example, 1238 is an increasing subsequence of S . Decreasing subsequences are defined similarly; 641 is a decreasing subsequence of S .
(a) List all the maximum-length increasing subsequences of S, and all the maximumlength decreasing subsequences.
Now let A be the set of numbers in S . (So A is the integers Œ1::9 for the example
above.) There are two straightforward linear orders for A. The first is numerical
order where A is ordered by the < relation. The second is to order the elements by
which comes first in S ; call this order <S . So for the example above, we would
have
6 <S 4 <S 7 <S 9 <S 1 <S 2 <S 5 <S 3 <S 8
Let be the product relation of the linear orders <s and <. That is, is defined
by the rule
a a0 WWD a < a0 AND a <S a0 :
So is a partial order on A (Section 10.9).
(b) Draw a diagram of the partial order on A. What are the maximal and minimal elements?
(c) Explain the connection between increasing and decreasing subsequences of S ,
and chains and anti-chains under .
(d) Prove that every sequence S of length n has an increasing subsequence of
p
p
length greater than n or a decreasing subsequence of length at least n.
Problems for Section 10.10
Practice Problems
Problem 10.54.
For each of the following relations, decide whether it is reflexive, whether it is
symmetric, whether it is transitive, and whether it is an equivalence relation.
(a) f.a; b/ j a and b are the same ageg
(b) f.a; b/ j a and b have the same parentsg
(c) f.a; b/ j a and b speak a common languageg
“mcs” — 2017/6/5 — 19:42 — page 436 — #444
436
Chapter 10
Directed graphs & Partial Orders
Problem 10.55.
For each of the binary relations below, state whether it is a strict partial order, a
weak partial order, an equivalence relation, or none of these. If it is a partial order,
state whether it is a linear order. If it is none, indicate which of the axioms for
partial-order and equivalence relations it violates.
(a) The superset relation on the power set pow f1; 2; 3; 4; 5g.
(b) The relation between any two nonnegative integers a and b such that a b
.mod 8/.
(c) The relation between propositional formulas G and H such that ŒG IMPLIES
H  is valid.
(d) The relation between propositional formulas G and H such that ŒG IFF H  is
valid.
(e) The relation ‘beats’ on Rock, Paper, and Scissors (for those who don’t know
the game Rock, Paper, Scissors, Rock beats Scissors, Scissors beats Paper, and
Paper beats Rock).
(f) The empty relation on the set of real numbers.
(g) The identity relation on the set of integers.
(h) The divisibility relation on the integers Z.
Class Problems
Problem 10.56.
Prove Theorem 10.10.4: The equivalence classes of an equivalence relation form a
partition of the domain.
Namely, let R be an equivalence relation on a set A and define the equivalence
class of an element a 2 A to be
ŒaR WWD fb 2 A j a R bg:
That is, ŒaR D R.a/.
(a) Prove that every block is nonempty and every element of A is in some block.
(b) Prove that if ŒaR \ ŒbR ¤ ;, then a R b. Conclude that the sets ŒaR for
a 2 A are a partition of A.
(c) Prove that a R b iff ŒaR D ŒbR .
“mcs” — 2017/6/5 — 19:42 — page 437 — #445
10.12. References
437
Problem 10.57.
For any total function f W A ! B define a relation f by the rule:
a f a0
iff f .a/ D f .a0 /:
(10.17)
(a) Sketch a proof that f is an equivalence relation on A.
(b) Prove that every equivalence relation R on a set A is equal to f for the
function f W A ! pow.A/ defined as
f .a/ WWD fa0 2 A j a R a0 g:
That is, f .a/ D R.a/.
Problem 10.58.
Let R be a binary relation on a set D. Each of the following formulas expresses the
fact that R has a familiar relational property such as reflexivity, asymmetry, transitivity. Predicate formulas have roman numerals i.,ii.,. . . , and relational formulas
(equalities and containments) are labelled with letters (a),(b),. . . .
Next to each of the relational formulas, write the roman numerals of all the predicate formulas equivalent to it. It is not necessary to name the property expressed,
but you can get partial credit if you do. For example, part (a) gets the label “i.” It
expresses irreflexivity.
i. 8d:
NOT.d
ii. 8d:
dRd
R d/
iii. 8c; d:
c R d IFF d R c
iv. 8c; d:
c R d IMPLIES d R c
v. 8c; d:
c R d IMPLIES NOT.d R c/
vi. 8c ¤ d:
c R d IMPLIES NOT.d R c/
vii. 8c ¤ d:
c R d IFF NOT.d R c/
viii. 8b; c; d:
.b R c AND c R d / IMPLIES b R d
ix. 8b; d:
Œ9c: .b R c AND c R d / IMPLIES b R d
x. 8b; d:
b R d IMPLIES Œ9c: .b R c AND c R d /
“mcs” — 2017/6/5 — 19:42 — page 438 — #446
438
Chapter 10
Directed graphs & Partial Orders
(a) R \ IdD D ;
(b) R R
1
(c) R D R
1
i.
(d) IdD R
(e) R ı R R
(f) R R ı R
(g) R \ R
1
(h) R R
1
IdD
(i) R \ IdR D R
(j) R \ R
1
1
\ IdR
D;
Homework Problems
Problem 10.59.
Let R1 and R2 be two equivalence relations on a set A. Prove or give a counterexample to the claims that the following are also equivalence relations:
(a) R1 \ R2 .
(b) R1 [ R2 .
Problem 10.60.
Say that vertices u; v in a digraph G are mutually connected and write
u
! v;
when there is a path from u to v and also a path from v to u.
(a) Prove that
! is an equivalence relation on V .G/.
(b) The blocks of the equivalence relation ! are called the strongly connected
components of G. Define a relation
on the strongly connected components of
G by the rule
C
Prove that
D
IFF
there is a path from some vertex in C to some vertex in D:
is a weak partial order on the strongly connected components.
“mcs” — 2017/6/5 — 19:42 — page 439 — #447
10.12. References
439
Exam Problems
Problem 10.61.
Let A be a nonempty set.
(a) Describe a single relation on A that is both an equivalence relation and a weak
partial order on A.
(b) Prove that the relation of part (a) is the only relation on A with these properties.
“mcs” — 2017/6/5 — 19:42 — page 440 — #448
“mcs” — 2017/6/5 — 19:42 — page 441 — #449
11
Communication Networks
Modeling communication networks is an important application of digraphs in computer science. In this such models, vertices represent computers, processors, and
switches; edges will represent wires, fiber, or other transmission lines through
which data flows. For some communication networks, like the internet, the corresponding graph is enormous and largely chaotic. Highly structured networks, by
contrast, find application in telephone switching systems and the communication
hardware inside parallel computers. In this chapter, we’ll look at some of the nicest
and most commonly used structured networks.
11.1
Routing
The kinds of communication networks we consider aim to transmit packets of data
between computers, processors, telephones, or other devices. The term packet
refers to some roughly fixed-size quantity of data—256 bytes or 4096 bytes or
whatever.
11.1.1
Complete Binary Tree
Let’s start with a complete binary tree. Figure 11.1 is an example with 4 inputs and
4 outputs.
In this diagram, and many that follow, the squares represent terminals—sources
and destinations for packets of data. The circles represent switches, which direct
packets through the network. A switch receives packets on incoming edges and
relays them forward along the outgoing edges. Thus, you can imagine a data
packet hopping through the network from an input terminal, through a sequence
of switches joined by directed edges, to an output terminal.
In a tree there is a unique path between every pair of vertices, so there is only
one way to route a packet of data from an input terminal to an output. For example,
the route of a packet traveling from input 1 to output 3 is shown in bold.
11.1.2
Routing Problems
Communication networks are supposed to get packets from inputs to outputs, with
each packet entering the network at its own input switch and arriving at its own
output switch. We’re going to consider several different communication network
“mcs” — 2017/6/5 — 19:42 — page 442 — #450
442
Chapter 11
Communication Networks
IN
0
OUT
0
IN
1
OUT
1
IN
2
OUT
2
IN
3
OUT
3
Figure 11.1 Binary Tree net with 4 inputs and outputs
designs, where each network has N inputs and N outputs; for convenience, we’ll
assume N is a power of two.
Which input is supposed to go where is specified by a permutation of Œ0::N 1.
So a permutation defines a routing problem: get a packet that starts at input i to
output .i /. A routing that solves a routing problem is a set P of paths from
each input to its specified output. That is, P is a set of paths Pi where Pi goes
from input i to output .i / for i 2 Œ0::N 1.
11.2
Routing Measures
11.2.1
Network Diameter
The delay between the time that a packets arrives at an input and arrives at its
designated output is a critical issue in communication networks. Generally, this
delay is proportional to the length of the path a packet follows. Assuming it takes
one time unit to travel across a wire, the delay of a packet will be the number of
wires it crosses going from input to output.
Packets are usually routed from input to output by the shortest path possible.
With a shortest-path routing, the worst-case delay is the distance between the input
and output that are farthest apart. This is called the diameter of the network. In
other words, the diameter of a network1 is the maximum length of any shortest
1 The
usual definition of diameter for a general graph (simple or directed) is the largest distance
between any two vertices, but in the context of a communication network we’re only interested in the
“mcs” — 2017/6/5 — 19:42 — page 443 — #451
11.2. Routing Measures
443
path between an input and an output. For example, in the complete binary tree
above, the distance from input 1 to output 3 is six. No input and output are farther
apart than this, so the diameter of this tree is also six.
More broadly, the diameter of a complete binary tree with N inputs and outputs
is 2 log N C2. This is quite good, because the logarithm function grows very slowly.
We could connect up 210 D 1024 inputs and outputs using a complete binary tree
and the worst input-output delay for any packet would be 2 log.210 / C 2 D 22.
Switch Size
One way to reduce the diameter of a network is to use larger switches. For example,
in the complete binary tree, most of the switches have three incoming edges and
three outgoing edges, which makes them 3 3 switches. If we had 4 4 switches,
then we could construct a complete ternary tree with an even smaller diameter. In
principle, we could even connect up all the inputs and outputs via a single monster
N N switch.
Of course this isn’t very productive. Using an N N switch would just conceal
the original network design problem inside this abstract switch. Eventually, we’ll
have to design the internals of the monster switch using simpler components, and
then we’re right back where we started. So, the challenge in designing a communication network is figuring out how to get the functionality of an N N switch
using fixed size, elementary devices, like 3 3 switches.
11.2.2
Switch Count
Another goal in designing a communication network is to use as few switches as
possible. In a complete binary tree, there is one “root” switch at the top, and the
number of switches doubles at successive rows, so the number of switches in an
N -input complete binary tree is 1 C 2 C 4 C 8 C C N . So the total number of
switches is 2N 1 by the formula for geometric sums (Problem 5.4). This is nearly
the best possible with 3 3 switches.
11.2.3
Network Latency
We’ll sometimes be choosing routings through a network that optimize some quantity besides delay. For example, in the next section we’ll be trying to minimize
packet congestion. When we’re not minimizing delay, shortest routings are not always the best, and in general, the delay of a packet will depend on how it is routed.
For any routing, the most delayed packet will be the one that follows the longest
path in the routing. The length of the longest path in a routing is called its latency.
distance between inputs and outputs, not between arbitrary pairs of vertices.
“mcs” — 2017/6/5 — 19:42 — page 444 — #452
444
Chapter 11
IN
Communication Networks
0
OUT
0
IN
1
OUT
1
IN
2
OUT
2
IN
3
OUT
3
IN
0
OUT
0
IN
1
OUT
1
IN
2
OUT
2
IN
3
OUT
3
Figure 11.2 Two Routings in the Binary Tree Net
The latency of a network depends on what’s being optimized. It is measured by
assuming that optimal routings are always chosen in getting inputs to their specified
outputs. That is, for each routing problem , we choose an optimal routing that
solves . Then network latency is defined to be the largest routing latency among
these optimal routings. Network latency will equal network diameter if routings
are always chosen to optimize delay, but it may be significantly larger if routings
are chosen to optimize something else.
For the networks we consider below, paths from input to output are uniquely
determined (in the case of the tree) or all paths are the same length, so network
latency will always equal network diameter.
11.2.4
Congestion
The complete binary tree has a fatal drawback: the root switch is a bottleneck. At
best, this switch must handle right and vice-versa. Passing all these packets through
a single switch could take a long time. At worst, if this switch fails, the network is
broken into two equal-sized pieces.
It’s true that if the routing problem is given by the identity permutation, Id.i / WWD
i , then there is an easy routing P that solves the problem: let Pi be the path from
input i up through one switch and back down to output i . On the other hand, if the
problem was given by .i / WWD .N 1/ i , then in any solution Q for , each path
Qi beginning at input i must eventually loop all the way up through the root switch
and then travel back down to output .N 1/ i . These two situations are illustrated
in Figure 11.2. We can distinguish between a “good” set of paths and a “bad” set
based on congestion. The congestion of a routing P is equal to the largest number
of paths in P that pass through a single switch. For example, the congestion of the
routing on the left is 1, since at most 1 path passes through each switch. However,
the congestion of the routing on the right is 4, since 4 paths pass through the root
switch (and the two switches directly below the root). Generally, lower congestion
is better since packets can be delayed at an overloaded switch.
“mcs” — 2017/6/5 — 19:42 — page 445 — #453
11.3. Network Designs
445
in0
in1
in2
in3
out0
out1
out2
out3
Figure 11.3 Two-dimensional Array with N D 4.
By extending the notion of congestion to networks, we can also distinguish between “good” and “bad” networks with respect to bottleneck problems. For each
routing problem for the network, we assume a routing is chosen that optimizes
congestion, that is, that has the minimum congestion among all routings that solve
. Then the largest congestion that will ever be suffered by a switch will be the
maximum congestion among these optimal routings. This “maximin” congestion
is called the congestion of the network.
So for the complete binary tree, the worst permutation would be .i / WWD .N
1/ i . Then in every possible solution for , every packet would have to follow
a path passing through the root switch. Thus, the max congestion of the complete
binary tree is N —which is horrible!
Let’s tally the results of our analysis so far:
network diameter switch size # switches congestion
complete binary tree 2 log N C 2
33
2N 1
N
11.3
Network Designs
11.3.1
2-D Array
Communication networks can also be designed as2-dimensional arrays or grids. A
2-D array with four inputs and outputs is shown in Figure 11.3.
The diameter in this example is 8, which is the number of edges between input 0
“mcs” — 2017/6/5 — 19:42 — page 446 — #454
446
Chapter 11
Communication Networks
and output 3. More generally, the diameter of an array with N inputs and outputs is
2N , which is much worse than the diameter of 2 log N C 2 in the complete binary
tree. But we get something in exchange: replacing a complete binary tree with an
array almost eliminates congestion.
Theorem 11.3.1. The congestion of an N -input array is 2.
Proof. First, we show that the congestion is at most 2. Let be any permutation.
Define a solution P for to be the set of paths, Pi , where Pi goes to the right from
input i to column .i / and then goes down to output .i /. Thus, the switch in row
i and column j transmits at most two packets: the packet originating at input i and
the packet destined for output j .
Next, we show that the congestion is at least 2. This follows because in any
routing problem , where .0/ D 0 and .N 1/ D N 1, two packets must
pass through the lower left switch.
As with the tree, the network latency when minimizing congestion is the same
as the diameter. That’s because all the paths between a given input and output are
the same length.
Now we can record the characteristics of the 2-D array.
network diameter switch size # switches congestion
complete binary tree 2 log N C 2
33
2N 1
N
2-D array
2N
22
N2
2
The crucial entry here is the number of switches, which is N 2 . This is a major
defect of the 2-D array; a network of size N D 1000 would require a million
2 2 switches! Still, for applications where N is small, the simplicity and low
congestion of the array make it an attractive choice.
11.3.2
Butterfly
The Holy Grail of switching networks would combine the best properties of the
complete binary tree (low diameter, few switches) and of the array (low congestion). The butterfly is a widely-used compromise between the two.
A good way to understand butterfly networks is as a recursive data type. The
recursive definition works better if we define just the switches and their connections, omitting the terminals. So we recursively define Fn to be the switches and
connections of the butterfly net with N WWD 2n input and output switches.
The base case is F1 with 2 input switches and 2 output switches connected as in
Figure 11.4.
“mcs” — 2017/6/5 — 19:42 — page 447 — #455
447
À
2 inputs
À
11.3. Network Designs
2 outputs
ND21
Figure 11.4 F1 , the Butterfly Net with N D 21 .
2n
2n
⎧
⎨
⎩
⎧
⎨
⎩
new inputs
Fn
n 1 outputs
2n+1
t t
Fn
Fn+1
Figure 11.5 Butterfly Net FnC1 with 2nC1 inputs from two Fn ’s.
In the constructor step, we construct FnC1 out of two Fn nets connected to a
new set of 2nC1 input switches, as shown in as in Figure 11.5. That is, the i th
and 2n C i th new input switches are each connected to the same two switches, the
i th input switches of each of two Fn components for i D 1; : : : ; 2n . The output
switches of FnC1 are simply the output switches of each of the Fn copies.
So FnC1 is laid out in columns of height 2nC1 by adding one more column of
switches to the columns in Fn . Since the construction starts with two columns
when n D 1, the FnC1 switches are arrayed in n C 1 columns. The total number of
switches is the height of the columns times the number 2nC1 .n C 1/ of columns.
Remembering that n D log N , we conclude that the Butterfly Net with N inputs
has N.log N C 1/ switches.
“mcs” — 2017/6/5 — 19:42 — page 448 — #456
448
Chapter 11
Communication Networks
Since every path in FnC1 from an input switch to an output is length-n C 1 the
diameter of the Butterfly net with 2nC1 inputs is this length plus two because of
the two edges connecting to the terminals (square boxes)—one edge from input
terminal to input switch (circle) and one from output switch to output terminal.
There is an easy recursive procedure to route a packet through the Butterfly Net.
In the base case, there is only one way to route a packet from one of the two inputs
to one of the two outputs. Now suppose we want to route a packet from an input
switch to an output switch in FnC1 . If the output switch is in the “top” copy of Fn ,
then the first step in the route must be from the input switch to the unique switch
it is connected to in the top copy; the rest of the route is determined by recursively
routing the rest of the way in the top copy of Fn . Likewise, if the output switch is in
the “bottom” copy of Fn , then the first step in the route must be to the switch in the
bottom copy, and the rest of the route is determined by recursively routing in the
bottom copy of Fn . In fact, this argument shows that the routing is unique: there is
exactly one path in the Butterfly Net from each input to each output, which implies
that the network latency when minimizing congestion
p is the same as the diameter.
The congestion
of
the
butterfly
network
is
about
N . More precisely, the conp
p
gestion is N if N is an even power of 2 and N=2 if N is an odd power of 2. A
simple proof of this appears in Problem 11.8.
Let’s add the butterfly data to our comparison table:
network diameter switch size
# switches
congestion
complete binary tree 2 log N C 2
33
2N 1
N
2-D array
2N
22
N2
2p
p
butterfly log N C 2
22
N.log.N / C 1/
N or N=2
The butterfly has lower congestion than the complete binary tree. It also uses fewer
switches and has lower diameter than the array. However, the butterfly does not
capture the best qualities of each network, but rather is a compromise somewhere
between the two. Our quest for the Holy Grail of routing networks goes on.
11.3.3
Beneš Network
In the 1960’s, a researcher at Bell Labs named Václav E. Beneš had a remarkable
idea. He obtained a marvelous communication network with congestion 1 by placing two butterflies back-to-back. This amounts to recursively growing Beneš nets
by adding both inputs and outputs at each stage. Now we recursively define Bn
to be the switches and connections (without the terminals) of the Beneš net with
N WWD 2n input and output switches.
The base case B1 with 2 input switches and 2 output switches is exactly the same
as F1 in Figure 11.4.
“mcs” — 2017/6/5 — 19:42 — page 449 — #457
À À
11.3. Network Designs
2n
2n
new inputs
449
Bn
2nC1
Bn
BnC1
new outputs
Figure 11.6 Beneš Net BnC1 with 2nC1 inputs from two Bn ’s.
In the constructor step, we construct BnC1 out of two Bn nets connected to a
new set of 2nC1 input switches and also a new set of 2nC1 output switches. This is
illustrated in Figure 11.6.
The i th and 2n C i th new input switches are each connected to the same two
switches: the i th input switches of each of two Bn components for i D 1; : : : ; 2n ,
exactly as in the Butterfly net. In addition, the i th and 2n C i th new output switches
are connected to the same two switches, namely, to the i th output switches of each
of two Bn components.
Now, BnC1 is laid out in columns of height 2nC1 by adding two more columns
of switches to the columns in Bn . So, the BnC1 switches are arrayed in 2.n C 1/
columns. The total number of switches is the number of columns times the height
2.n C 1/2nC1 of the columns.
All paths in BnC1 from an input switch to an output are length 2.n C 1/ 1, and
the diameter of the Beneš net with 2nC1 inputs is this length plus two because of
the two edges connecting to the terminals.
So Beneš has doubled the number of switches and the diameter, but by doing so
he has completely eliminated congestion problems! The proof of this fact relies on
a clever induction argument that we’ll come to in a moment. Let’s first see how the
“mcs” — 2017/6/5 — 19:42 — page 450 — #458
450
Chapter 11
Communication Networks
Beneš network stacks up:
# switches
congestion
network diameter switch size
complete binary tree 2 log N C 2
33
2N 1
N
2N
22
N2
2p
2-D array
p
butterfly log N C 2
22
N.log.N / C 1/
N or N=2
22
2N log N
1
Beneš 2 log N C 1
The Beneš network has small size and diameter, and it completely eliminates congestion. The Holy Grail of routing networks is in hand!
Theorem 11.3.2. The congestion of the N -input Beneš network is 1.
Proof. By induction on n where N D 2n . So the induction hypothesis is
P .n/ WWD the congestion of Bn is 1:
Base case (n D 1): B1 D F1 is shown in Figure 11.4. The unique routings in F1
have congestion 1.
Inductive step: We assume that the congestion of an N D 2n -input Beneš network
is 1 and prove that the congestion of a 2N -input Beneš network is also 1.
Digression. Time out! Let’s work through an example, develop some intuition,
and then complete the proof. In the Beneš network shown in Figure 11.7 with
N D 8 inputs and outputs, the two 4-input/output subnetworks are in dashed boxes.
By the inductive assumption, the subnetworks can each route an arbitrary permutation with congestion 1. So if we can guide packets safely through just the first
and last levels, then we can rely on induction for the rest! Let’s see how this works
in an example. Consider the following permutation routing problem:
.0/ D 1
.4/ D 3
.1/ D 5
.5/ D 6
.2/ D 4
.6/ D 0
.3/ D 7
.7/ D 2
We can route each packet to its destination through either the upper subnetwork
or the lower subnetwork. However, the choice for one packet may constrain the
choice for another. For example, we cannot route both packet 0 and packet 4
through the same network, since that would cause two packets to collide at a single switch, resulting in congestion. Rather, one packet must go through the upper
network and the other through the lower network. Similarly, packets 1 and 5, 2 and
“mcs” — 2017/6/5 — 19:42 — page 451 — #459
11.3. Network Designs
451
in0
out0
in1
out1
in2
out2
in3
out3
in4
out4
in5
out5
in6
out6
in7
out7
Figure 11.7 Beneš net B3 .
6, and 3 and 7 must be routed through different networks. Let’s record these constraints in a graph. The vertices are the 8 packets. If two packets must pass through
different networks, then there is an edge between them. Thus, our constraint graph
looks like this:
1
5
2
0
6
4
7
3
Notice that at most one edge is incident to each vertex.
The output side of the network imposes some further constraints. For example,
the packet destined for output 0 (which is packet 6) and the packet destined for
output 4 (which is packet 2) cannot both pass through the same network; that would
require both packets to arrive from the same switch. Similarly, the packets destined
for outputs 1 and 5, 2 and 6, and 3 and 7 must also pass through different switches.
We can record these additional constraints in our graph with gray edges:
“mcs” — 2017/6/5 — 19:42 — page 452 — #460
452
Chapter 11
Communication Networks
1
5
2
0
6
4
7
3
Notice that at most one new edge is incident to each vertex. The two lines drawn
between vertices 2 and 6 reflect the two different reasons why these packets must
be routed through different networks. However, we intend this to be a simple graph;
the two lines still signify a single edge.
Now here’s the key insight: suppose that we could color each vertex either red
or blue so that adjacent vertices are colored differently. Then all constraints are
satisfied if we send the red packets through the upper network and the blue packets
through the lower network. Such a 2-coloring of the graph corresponds to a solution to the routing problem. The only remaining question is whether the constraint
graph is 2-colorable, which is easy to verify:
Lemma 11.3.3. Prove that if the edges of a graph can be grouped into two sets such
that every vertex has at most 1 edge from each set incident to it, then the graph is
2-colorable.
Proof. It is not hard to show that a graph is 2-colorable iff every cycle in it has even
length (see Theorem 12.8.3). We’ll take this for granted here.
So all we have to do is show that every cycle has even length. Since the two sets
of edges may overlap, let’s call an edge that is in both sets a doubled edge.
There are two cases:
Case 1: [The cycle contains a doubled edge.] No other edge can be incident
to either of the endpoints of a doubled edge, since that endpoint would then be
incident to two edges from the same set. So a cycle traversing a doubled edge has
nowhere to go but back and forth along the edge an even number of times.
Case 2: [No edge on the cycle is doubled.] Since each vertex is incident to
at most one edge from each set, any path with no doubled edges must traverse
successive edges that alternate from one set to the other. In particular, a cycle must
traverse a path of alternating edges that begins and ends with edges from different
sets. This means the cycle has to be of even length.
For example, here is a 2-coloring of the constraint graph:
“mcs” — 2017/6/5 — 19:42 — page 453 — #461
11.3. Network Designs
453
blue
1
red
5
red 0
2 red
blue 4
6 blue
7
blue
3
red
The solution to this graph-coloring problem provides a start on the packet routing
problem:
We can complete the routing in the two smaller Beneš networks by induction!
Back to the proof. End of Digression.
Let be an arbitrary permutation of Œ0::N
1. Let G be the graph whose
vertices are packet numbers 0; 1; : : : ; N 1 and whose edges come from the union
of these two sets:
E1 WWDfhu—vi j ju
vj D N=2g; and
E2 WWDfhu—wi j j.u/
.w/j D N=2g:
Now any vertex u is incident to at most two edges: a unique edge hu—vi 2 E1
and a unique edge hu—wi 2 E2 . So according to Lemma 11.3.3, there is a 2coloring for the vertices of G. Now route packets of one color through the upper
subnetwork and packets of the other color through the lower subnetwork. Since
for each edge in E1 , one vertex goes to the upper subnetwork and the other to the
lower subnetwork, there will not be any conflicts in the first level. Since for each
edge in E2 , one vertex comes from the upper subnetwork and the other from the
lower subnetwork, there will not be any conflicts in the last level. We can complete
the routing within each subnetwork by the induction hypothesis P .n/.
Problems for Section 11.2
Exam Problems
Problem 11.1.
Consider the following communication network:
(a) What is the max congestion?
(b) Give an input/output permutation 0 that forces maximum congestion.
(c) Give an input/output permutation 1 that allows minimum congestion.
“mcs” — 2017/6/5 — 19:42 — page 454 — #462
454
Chapter 11
Communication Networks
in0
in1
out0
in2
out1
out2
(d) What is the latency for the permutation 1 ? (If you could not find 1 , just
choose a permutation and find its latency.)
Problems for Section 11.3
Class Problems
Problem 11.2.
The Beneš network has a max congestion of one—every permutation can be routed
in such a way that a single packet passes through each switch. Let’s work through
an example. A diagram of the Beneš network B3 of size N D 8 appears in Figure 11.7. The two subnetworks of size N D 4 are marked. We’ll refer to these as
the upper and lower subnetworks.
(a) Now consider the following permutation routing problem:
.0/ D 3
.4/ D 2
.1/ D 1
.5/ D 0
.2/ D 6
.6/ D 7
.3/ D 5
.7/ D 4
Each packet must be routed through either the upper subnetwork or the lower subnetwork. Construct a graph with vertices numbered by integers 0 to 7 and draw a
dashed edge between each pair of packets that cannot go through the same subnetwork because a collision would occur in the second column of switches.
(b) Add a solid edge in your graph between each pair of packets that cannot go
through the same subnetwork because a collision would occur in the next-to-last
column of switches.
“mcs” — 2017/6/5 — 19:42 — page 455 — #463
11.3. Network Designs
455
(c) Assign colors red and blue to the vertices of your graph so that vertices that
are adjacent by either a dashed or a solid edge get different colors. Why must this
be possible, regardless of the permutation ?
(d) Suppose that red vertices correspond to packets routed through the upper subnetwork and blue vertices correspond to packets routed through the lower subnetwork. Referring to the Beneš network shown in Figure 11.6, indicate the first and
last edge traversed by each packet.
(e) All that remains is to route packets through the upper and lower subnetworks.
One way to do this is by applying the procedure described above recursively on
each subnetwork. However, since the remaining problems are small, see if you can
complete all the paths on your own.
Problem 11.3.
A multiple binary-tree network has N inputs and N outputs, where N is a power
of 2. Each input is connected to the root of a binary tree with N=2 leaves and with
edges pointing away from the root. Likewise, each output is connected to the root
of a binary tree with N=2 leaves and with edges pointing toward the root.
Two edges point from each leaf of an input tree, and each of these edges points
to a leaf of an output tree. The matching of leaf edges is arranged so that for every
input and output tree, there is an edge from a leaf of the input tree to a leaf of the
output tree, and every output tree leaf has exactly two edges pointing to it.
(a) Draw such a multiple binary-tree net for N D 4.
(b) Fill in the table, and explain your entries.
# switches switch size diameter max congestion
Problem 11.4.
The n-input 2-D array network was shown to have congestion 2. An n-input 2layer array consisting of two n-input 2-D Arrays connected as pictured below for
n D 4.
In general, an n-input 2-layer array has two layers of switches, with each layer
connected like an n-input 2-D array. There is also an edge from each switch in the
first layer to the corresponding switch in the second layer. The inputs of the 2-layer
“mcs” — 2017/6/5 — 19:42 — page 456 — #464
456
Chapter 11
Communication Networks
in0
in1
in2
in3
out0
out1
out2
out3
array enter the left side of the first layer, and the n outputs leave from the bottom
row of either layer.
(a) For any given input-output permutation, there is a way to route packets that
achieves congestion 1. Describe how to route the packets in this way.
(b) What is the latency of a routing designed to minimize latency?
(c) Explain why the congestion of any minimum latency (CML) routing of packets
through this network is greater than the network’s congestion.
Problem 11.5.
A 5-path communication network is shown below. From this, it’s easy to see what
an n-path network would be. Fill in the table of properties below, and be prepared
to justify your answers.
network # switches switch size diameter max congestion
5-path
n-path
Problem 11.6.
“mcs” — 2017/6/5 — 19:42 — page 457 — #465
11.3. Network Designs
457
in0
in1
in2
in3
in4
out0
out1
out2
out3
out4
Figure 11.8 5-Path
Tired of being a TA, Megumi has decided to become famous by coming up with a
new, better communication network design. Her network has the following specifications: every input node will be sent to a butterfly network, a Beneš network and
a 2-d array network. At the end, the outputs of all three networks will converge on
the new output.
In the Megumi-net a minimum latency routing does not have minimum congestion. The latency for min-congestion (LMC) of a net is the best bound on latency
achievable using routings that minimize congestion. Likewise, the congestion for
min-latency (CML) is the best bound on congestion achievable using routings that
minimize latency.
out1
out2
in1
in2
Butterfly
out3
in3
.
.
.
Beneš
.
.
inN
.
2-d Array
outN
Fill in the following chart for Megumi’s new net and explain your answers.
“mcs” — 2017/6/5 — 19:42 — page 458 — #466
458
Chapter 11
Communication Networks
network
Megumi’s net
diameter
# switches
congestion
LMC
Homework Problems
Problem 11.7.
Louis Reasoner figures that, wonderful as the Beneš network may be, the butterfly
network has a few advantages, namely: fewer switches, smaller diameter, and an
easy way to route packets through it. So Louis designs an N -input/output network
he modestly calls a Reasoner-net with the aim of combining the best features of
both the butterfly and Beneš nets:
The i th input switch in a Reasoner-net connects to two switches, ai and
bi , and likewise, the j th output switch has two switches, yj and zj ,
connected to it. Then the Reasoner-net has an N -input Beneš network
connected using the ai switches as input switches and the yj switches
as its output switches. The Reasoner-net also has an N -input butterfly
net connected using the bi switches as inputs and¡ the zj switches as
outputs.
In the Reasoner-net a minimum latency routing does not have minimum congestion. The latency for min-congestion (LMC) of a net is the best bound on latency
achievable using routings that minimize congestion. Likewise, the congestion for
min-latency (CML) is the best bound on congestion achievable using routings that
minimize latency.
Fill in the following chart for the Reasoner-net and briefly explain your answers.
diameter switch size(s) # switches congestion LMC CML
Problem 11.8.
p
Show that the congestion of the butterfly net, Fn , is exactly N when n is even.
Hint:
There is a unique path from each input to each output, so the congestion is
the maximum number of messages passing through a vertex for any routing
problem.
CML
“mcs” — 2017/6/5 — 19:42 — page 459 — #467
11.3. Network Designs
459
If v is a vertex in column i of the butterfly network, there is a path from exactly 2i input vertices to v and a path from v to exactly 2n i output vertices.
At which column of the butterfly network must the congestion be worst?
What is the congestion of the topmost switch in that column of the network?
“mcs” — 2017/6/5 — 19:42 — page 460 — #468
“mcs” — 2017/6/5 — 19:42 — page 461 — #469
12
Simple Graphs
Simple graphs model relationships that are symmetric, meaning that the relationship
is mutual. Examples of such mutual relationships are being married, speaking the
same language, not speaking the same language, occurring during overlapping time
intervals, or being connected by a conducting wire. They come up in all sorts of
applications, including scheduling, constraint satisfaction, computer graphics, and
communications, but we’ll start with an application designed to get your attention:
we are going to make a professional inquiry into sexual behavior. Specifically,
we’ll look at some data about who, on average, has more opposite-gender partners:
men or women.
Sexual demographics have been the subject of many studies. In one of the largest,
researchers from the University of Chicago interviewed a random sample of 2500
people over several years to try to get an answer to this question. Their study,
published in 1994 and entitled The Social Organization of Sexuality, found that
men have on average 74% more opposite-gender partners than women.
Other studies have found that the disparity is even larger. In particular, ABC
News claimed that the average man has 20 partners over his lifetime, and the average woman has 6, for a percentage disparity of 233%. The ABC News study,
aired on Primetime Live in 2004, purported to be one of the most scientific ever
done, with only a 2.5% margin of error. It was called “American Sex Survey: A
peek between the sheets”—raising some questions about the seriousness of their
reporting.
Yet again in August, 2007, the New York Times reported on a study by the
National Center for Health Statistics of the U.S. government showing that men had
seven partners while women had four. So, whose numbers do you think are more
accurate: the University of Chicago, ABC News, or the National Center?
Don’t answer—this is a trick question designed to trip you up. Using a little
graph theory, we’ll explain why none of these findings can be anywhere near the
truth.
12.1
Vertex Adjacency and Degrees
Simple graphs are defined in almost the same way as digraphs, except that edges are
undirected—they connect two vertices without pointing in either direction between
the vertices. So instead of a directed edge hv ! wi which starts at vertex v and
“mcs” — 2017/6/5 — 19:42 — page 462 — #470
462
Chapter 12
Simple Graphs
ends at vertex w, a simple graph only has an undirected edge hv—wi that connects
v and w.
Definition 12.1.1. A simple graph G consists of a nonempty set, V .G/, called
the vertices of G, and a set E.G/ called the edgesAn element of V .G/ is called
a vertex. An element of E.G/ is an undirected edge or simply an “edge.” An
undirected edge has two vertices u ¤ v called its endpoints. Such an edge can be
represented by the two element set fu; vg. The notation hu—vi denotes this edge.
Both hu—vi and hv—ui describe the same undirected edge, whose endpoints
are u and v.
h
b
a
d
f
g
c
e
i
Figure 12.1 An example of a graph with 9 vertices and 8 edges.
For example, let H be the graph pictured in Figure 12.1. The vertices of H
correspond to the nine dots in Figure 12.1, that is,
V .H / D fa; b; c; d; e; f; g; h; i g :
The edges correspond to the eight lines, that is,
E.H / D f ha—bi ; ha—ci ; hb—d i ; hc—d i ; hc—ei ; he—f i ; he—gi ; hh—i i g:
Mathematically, that’s all there is to the graph H .
Definition 12.1.2. Two vertices in a simple graph are said to be adjacent iff they
are the endpoints of the same edge, and an edge is said to be incident to each of its
endpoints. The number of edges incident to a vertex v is called the degree of the
vertex and is denoted by deg.v/. Equivalently, the degree of a vertex is the number
of vertices adjacent to it.
For example, for the graph H of Figure 12.1, vertex a is adjacent to vertex b, and
b is adjacent to d . The edge ha—ci is incident to its endpoints a and c. Vertex h
has degree 1, d has degree 2, and deg.e/ D 3. It is possible for a vertex to have
degree 0, in which case it is not adjacent to any other vertices. A simple graph G
does not need to have any edges at all. jE.G/j could be zero, implying that the
“mcs” — 2017/6/5 — 19:42 — page 463 — #471
12.2. Sexual Demographics in America
463
degree of every vertex would also be zero. But a simple graph must have at least
one vertex—jV .G/j is required to be at least one.
Note that in a simple graph there cannot be more than one edge between the
same two vertices.1 Also self-loops—edges that begin and at the same vertex—are
not allowed in simple graphs.
For the rest of this chapter we’ll use “graphs” as an abbreviation for “simple
graphs.”
A synonym for “vertices” is “nodes,” and we’ll use these words interchangeably.
Simple graphs are sometimes called networks, edges are sometimes called arcs or
lines. We mention this as a “heads up” in case you look at other graph theory
literature; we won’t use these words.
12.2
Sexual Demographics in America
Let’s model the question of heterosexual partners in graph theoretic terms. To do
this, we’ll let G be the graph whose vertices V are all the people in America. Then
we split V into two separate subsets: M which contains all the males, and F which
contains all the females.2 We’ll put an edge between a male and a female iff they
have been sexual partners. This graph is pictured in Figure 12.2 with males on the
left and females on the right.
Actually, this is a pretty hard graph to figure out, let alone draw. The graph is
enormous: the US population is about 300 million, so jV j 300M . Of these,
approximately 50.8% are female and 49.2% are male, so jM j 147:6M , and
jF j 152:4M . And we don’t even have trustworthy estimates of how many edges
there are, let alone exactly which couples are adjacent. But it turns out that we
don’t need to know any of this—we just need to figure out the relationship between
the average number of partners per male and partners per female. To do this, we
note that every edge has exactly one endpoint at an M vertex (remember, we’re
only considering male-female relationships); so the sum of the degrees of the M
vertices equals the number of edges. For the same reason, the sum of the degrees
of the F vertices equals the number of edges. So these sums are equal:
X
X
deg.x/ D
deg.y/:
x2M
y2F
Now suppose we divide both sides of this equation by the product of the sizes of
1 Multigraphs that have more than one edge with the same two endpoints are sometimes convenient, but they are not needed for the applications we present.
2 For simplicity, we’ll ignore the possibility of someone being both a man and a woman, or neither.
“mcs” — 2017/6/5 — 19:42 — page 464 — #472
464
Chapter 12
Simple Graphs
M
F
Figure 12.2 The sex partners graph.
the two sets, jM j jF j:
P
x2M deg.x/
jM j
1
D
jF j
P
y2F
deg.y/
!
jF j
1
jM j
The terms above in parentheses are the average degree of an M vertex and the
average degree of an F vertex. So we know:
Avg. deg in M D
jF j
Avg. deg in F
jM j
(12.1)
In other words, we’ve proved that the average number of female partners of
males in the population compared to the average number of males per female is
determined solely by the relative number of males and females in the population.
Now the Census Bureau reports that there are slightly more females than males
in America; in particular jF j=jM j is about 1.035. So we know that males have
on average 3.5% more opposite-gender partners than females, and that this tells us
nothing about any sex’s promiscuity or selectivity. Rather, it just has to do with the
relative number of males and females. Collectively, males and females have the
same number of opposite gender partners, since it takes one of each set for every
partnership, but there are fewer males, so they have a higher ratio. This means that
the University of Chicago, ABC, and the Federal government studies are way off.
After a huge effort, they gave a totally wrong answer.
There’s no definite explanation for why such surveys are consistently wrong.
One hypothesis is that males exaggerate their number of partners—or maybe females downplay theirs—but these explanations are speculative. Interestingly, the
“mcs” — 2017/6/5 — 19:42 — page 465 — #473
12.3. Some Common Graphs
465
principal author of the National Center for Health Statistics study reported that she
knew the results had to be wrong, but that was the data collected, and her job was
to report it.
The same underlying issue has led to serious misinterpretations of other survey
data. For example, a couple of years ago, the Boston Globe ran a story on a survey
of the study habits of students on Boston area campuses. Their survey showed that
on average, minority students tended to study with non-minority students more than
the other way around. They went on at great length to explain why this “remarkable
phenomenon” might be true. But it’s not remarkable at all. Using our graph theory
formulation, we can see that all it says is that there are fewer students in a minority
than students not in that minority, which is, of course, what “minority” means.
12.2.1
Handshaking Lemma
The previous argument hinged on the connection between a sum of degrees and the
number of edges. There is a simple connection between these in any graph:
Lemma 12.2.1. The sum of the degrees of the vertices in a graph equals twice the
number of edges.
Proof. Every edge contributes two to the sum of the degrees, one for each of its
endpoints.
We refer to Lemma 12.2.1 as the Handshaking Lemma: if we total up the number
of people each person at a party shakes hands with, the total will be twice the
number of handshakes that occurred.
12.3
Some Common Graphs
Some graphs come up so frequently that they have names. A complete graph Kn
has n vertices and an edge between every two vertices, for a total of n.n 1/=2
edges. For example, K5 is shown in Figure 12.3.
The empty graph has no edges at all. For example, the five-vertex empty graph
is shown in Figure 12.4.
An n-vertex graph containing n 1 edges in sequence is known as
a line graph Ln . More formally, Ln has
V .Ln / D fv1 ; v2 ; : : : ; vn g
and
E.Ln / D f hv1 —v2 i ; hv2 —v3 i ; : : : ; hvn
1 —vn i g
“mcs” — 2017/6/5 — 19:42 — page 466 — #474
466
Chapter 12
Simple Graphs
Figure 12.3 K5 : the complete graph on five nodes.
Figure 12.4 An empty graph with five nodes.
For example, L5 is pictured in Figure 12.5.
There is also a one-way infinite line graph L1 which can be defined by letting
the nonnegative integers N be the vertices with edges hk—.k C 1/i for all k 2 N.
If we add the edge hvn —v1 i to the line graph Ln , we get a graph called a lengthn cycle Cn . Figure 12.6 shows a picture of length-5 cycle.
12.4
Isomorphism
Two graphs that look different might actually be the same in a formal sense. For
example, the two graphs in Figure 12.7 are both four-vertex, five-edge graphs and
Figure 12.5 L5 : a 5-vertex line graph.
“mcs” — 2017/6/5 — 19:42 — page 467 — #475
12.4. Isomorphism
467
Figure 12.6 C5 : a 5-node cycle graph.
a
d
(a)
b
1
c
4
2
3
(b)
Figure 12.7 Two Isomorphic graphs.
you get graph (b) by a 90o clockwise rotation of graph (a).
Strictly speaking, these graphs are different mathematical objects, but this difference doesn’t reflect the fact that the two graphs can be described by the same
picture—except for the labels on the vertices. This idea of having the same picture
“up to relabeling” can be captured neatly by adapting Definition 10.7.1 of isomorphism of digraphs to handle simple graphs. An isomorphism between two graphs
is an edge-preserving bijection between their sets of vertices:
Definition 12.4.1. An isomorphism between graphs G and H is a bijection f W
V .G/ ! V .H / such that
hu—vi 2 E.G/ IFF hf .u/—f .v/i 2 E.H /
for all u; v 2 V .G/. Two graphs are isomorphic when there is an isomorphism
between them.
Here is an isomorphism f between the two graphs in Figure 12.7:
f .a/ WWD 2
f .c/ WWD 4
f .b/ WWD 3
f .d / WWD 1:
You can check that there is an edge between two vertices in the graph on the left if
and only if there is an edge between the two corresponding vertices in the graph on
the right.
“mcs” — 2017/6/5 — 19:42 — page 468 — #476
468
Chapter 12
Simple Graphs
Figure 12.8 Isomorphic C5 graphs.
Two isomorphic graphs may be drawn very differently. For example, Figure 12.8
shows two different ways of drawing C5 .
Notice that if f is an isomorphism between G and H , then f 1 is an isomorphism between H and G. Isomorphism is also transitive because the composition
of isomorphisms is an isomorphism. In fact, isomorphism is an equivalence relation.
Isomorphism preserves the connection properties of a graph, abstracting out what
the vertices are called, what they are made out of, or where they appear in a drawing
of the graph. More precisely, a property of a graph is said to be preserved under
isomorphism if whenever G has that property, every graph isomorphic to G also
has that property. For example, since an isomorphism is a bijection between sets of
vertices, isomorphic graphs must have the same number of vertices. What’s more,
if f is a graph isomorphism that maps a vertex v of one graph to the vertex f .v/
of an isomorphic graph, then by definition of isomorphism, every vertex adjacent
to v in the first graph will be mapped by f to a vertex adjacent to f .v/ in the
isomorphic graph. Thus, v and f .v/ will have the same degree. If one graph has a
vertex of degree four and another does not, then they can’t be isomorphic. In fact,
they can’t be isomorphic if the number of degree-four vertices in each of the graphs
is not the same.
Looking for preserved properties can make it easy to determine that two graphs
are not isomorphic, or to guide the search for an isomorphism when there is one.
It’s generally easy in practice to decide whether two graphs are isomorphic. However, no one has yet found a procedure for determining whether two graphs are
isomorphic that is guaranteed to run in polynomial time on all pairs of graphs.3
Having such a procedure would be useful. For example, it would make it easy
to search for a particular molecule in a database given the molecular bonds. On
the other hand, knowing there is no such efficient procedure would also be valuable: secure protocols for encryption and remote authentication can be built on the
3 A procedure runs in polynomial time when it needs an amount of time of at most p.n/, where n
is the total number of vertices and p./ is a fixed polynomial.
“mcs” — 2017/6/5 — 19:42 — page 469 — #477
12.5. Bipartite Graphs & Matchings
469
hypothesis that graph isomorphism is computationally exhausting.
The definitions of bijection and isomorphism apply to infinite graphs as well as
finite graphs, as do most of the results in the rest of this chapter. But graph theory
focuses mostly on finite graphs, and we will too.
In the rest of this chapter we’ll assume graphs are finite.
We’ve actually been taking isomorphism for granted ever since we wrote “Kn
has n vertices. . . ” at the beginning of Section 12.3.
Graph theory is all about properties preserved by isomorphism.
12.5
Bipartite Graphs & Matchings
There were two kinds of vertices in the “Sex in America” graph, males and females,
and edges only went between the two kinds. Graphs like this come up so frequently
that they have earned a special name: bipartite graphs.
Definition 12.5.1. A bipartite graph is a graph whose vertices can be partitioned
into two blocks, L.G/ and R.G/, such that every edge has one endpoint in L.G/
and the other endpoint in R.G/.
So every bipartite graph looks something like the graph in Figure 12.2.
12.5.1
The Bipartite Matching Problem
The bipartite matching problem is related to the sex-in-America problem that we
just studied; only now, the goal is to get everyone happily married. As you might
imagine, this is not possible for a variety of reasons, not the least of which is the
fact that there are more women in America than men. So, it is simply not possible
to marry every woman to a man so that every man is married at most once.
But what about getting a mate for every man so that every woman is married at
most once? Is it possible to do this so that each man is paired with a woman that
he likes? The answer, of course, depends on the bipartite graph that represents who
likes who, but the good news is that it is possible to find natural properties of the
who-likes-who graph that completely determine the answer to this question.
In general, suppose that we have a set of men and an equal-sized or larger set of
women, and there is a graph with an edge between a man and a woman if the man
likes the woman. In this scenario, the “likes” relationship need not be symmetric,
since for the time being, we will only worry about finding a mate for each man
“mcs” — 2017/6/5 — 19:42 — page 470 — #478
470
Chapter 12
Simple Graphs
Alice
Chuck
Martha
Tom
Sara
Michael
Jane
John
Mergatroid
Figure 12.9 A graph where an edge between a man and woman denotes that the
man likes the woman.
that he likes.4 Later, we will consider the “likes” relationship from the female
perspective as well. For example, we might obtain the graph in Figure 12.9.
A matching is defined to be an assignment of a woman to each man so that
different men are assigned to different women, and a man is always assigned a
woman that he likes. For example, one possible matching for the men is shown in
Figure 12.10.
12.5.2
The Matching Condition
A famous result known as Hall’s Matching Theorem gives necessary and sufficient
conditions for the existence of a matching in a bipartite graph. It turns out to be a
remarkably useful mathematical tool.
We’ll state and prove Hall’s Theorem using man-likes-woman terminology. Define the set of women liked by a given set of men to consist of all women liked by
at least one of those men. For example, the set of women liked by Tom and John in
Figure 12.9 consists of Martha, Sara and Mergatroid. For us to have any chance at
all of matching up the men, the following matching condition must hold:
The Matching Condition: every subset of men likes at least as large a set of women.
For example, we cannot find a matching if some set of 4 men like only 3 women.
Hall’s Theorem says that this necessary condition is actually sufficient; if the matching condition holds, then a matching exists.
4 It’s convenient to describe the matching problem in terms of men and women, since it let’s us
say things like “a man’s wife” instead of “a left hand node’s matching right hand node.” We don’t
think marriage must be heterosexual, or that men should be given first choice of mate.
“mcs” — 2017/6/5 — 19:42 — page 471 — #479
12.5. Bipartite Graphs & Matchings
471
Alice
Chuck
Martha
Tom
Sara
Michael
Jane
John
Mergatroid
Figure 12.10 One possible matching for the men is shown with bold edges. For
example, John is matched with Mergatroid.
Theorem 12.5.2. A matching for a set M of men with a set W of women can be
found if and only if the matching condition holds.
Proof. First, let’s suppose that a matching exists and show that the matching condition holds. For any subset of men, each man likes at least the woman he is matched
with and a woman is matched with at most one man. Therefore, every subset of
men likes at least as large a set of women. Thus, the matching condition holds.
Next, let’s suppose that the matching condition holds and show that a matching
exists. We use strong induction on jM j, the number of men, on the predicate:
P .m/ WWD if the matching condition holds for a set, M ,
of m men, then there is a matching for M .
Base case (jM j D 1): If jM j D 1, then the matching condition implies that the
lone man likes at least one woman, and so a matching exists.
Inductive Step: Suppose that jM j D m C 1 2. To find a matching for M , there
are two cases.
Case 1: Every nonempty subset of at most m men likes a strictly larger set of
women. In this case, we have some latitude: we pair an arbitrary man with
a woman he likes and send them both away. This leaves m men and one
fewer women, and the matching condition will still hold. So the induction
hypothesis P .m/ implies we can match the remaining m men.
“mcs” — 2017/6/5 — 19:42 — page 472 — #480
472
Chapter 12
Simple Graphs
Case 2: Some nonempty subset X of at most m men likes an equal-size set Y of
women. The matching condition must hold within X, so the strong induction
hypothesis implies we can match the men in X with the women in Y . This
leaves the problem of matching the set M X of men to the set W Y of
women.
But the problem of matching M X against W Y also satisfies the Matching condition, because any subset of men in M X who liked fewer women
in W Y would imply there was a set of men who liked fewer women in the
whole set W . Namely, if a subset M0 M X liked only a strictly smaller
subset of women W0 W Y , then the set M0 [ X of men would like only
women in the strictly smaller set W0 [ Y . So again the strong induction hypothesis implies we can match the men in M X with the women in W Y ,
which completes a matching for M .
So in both cases, there is a matching for the men, which completes the proof of
the Inductive step. The theorem follows by induction.
The proof of Theorem 12.5.2 gives an algorithm for finding a matching in a bipartite graph, albeit not a very efficient one. However, efficient algorithms for finding a matching in a bipartite graph do exist. Thus, if a problem can be reduced to
finding a matching, instances of the problem can be solved in a reasonably efficient
way.
A Formal Statement
Let’s restate Theorem 12.5.2 in mathematical terms so that you’ll not always be
condemned to saying, “Now this group of men likes at least as many women. . . ”
Definition 12.5.3. A matching in a graph G is a set of edges M E.G/ such that
no two edges in M are incident to the same vertex. The vertices incident to an edge
in M are said to be covered by M . A matching is said to be perfect when it covers
V .G/. The set N.S / of neighbors of some set S V .G/ of vertices is the image
of S under the edge-relation, that is,
N.S/ WWD f r j hs—ri 2 E.G/ for some s 2 S g:
S is called a bottleneck if
jS j > j N.S /j:
Theorem 12.5.4 (Hall’s Theorem). Let G be a bipartite graph. There is a matching
in G that covers L.G/ iff no subset of L.G/ is a bottleneck.
“mcs” — 2017/6/5 — 19:42 — page 473 — #481
12.5. Bipartite Graphs & Matchings
473
An Easy Matching Condition
The bipartite matching condition requires that every subset of men has a certain
property. In general, verifying that every subset has some property, even if it’s easy
to check any particular subset for the property, quickly becomes overwhelming
because the number of subsets of even relatively small sets is enormous—over a
billion subsets for a set of size 30. However, there is a simple property of vertex
degrees in a bipartite graph that guarantees the existence of a matching. Call a
bipartite graph degree-constrained if vertex degrees on the left are at least as large
as those on the right. More precisely,
Definition 12.5.5. A bipartite graph G is degree-constrained when there is an integer d 1 such that
deg.l/ d deg.r/
(12.2)
for every l 2 L.G/ and r 2 R.G/.
For example, the graph in Figure 12.9 is degree-constrained with d D 2, since
every vertex on the left is adjacent to at least two vertices on the right while every
vertex on the right is adjacent to at most two vertices on the left.
Theorem 12.5.6. If G is a degree-constrained bipartite graph, then there is a
matching that covers L.G/.
Proof. Suppose G satisfies (12.2), and let S be a subset of L.G/. We will show
that S satisfies Hall’s condition, namely,
j N.S /j jS j:
(12.3)
Let DS be the set of edges with an endpoint in a S. According to (12.2), every
vertex in S is the endpoint of at least d edges,
jDS j d jS j:
Likewise, any edge in DS has an endpoint in N.S / by definition. According
to (12.2), every vertex in N.S / is incident to at most d edges, so we also have
d j N.S /j jDS j:
Combining the above inequalities gives
d j N.S /j d jS j:
Now cancelling d completes the derivation of equation (12.3).
A graph is regular when all its vertice have the same degree. Regular graphs
come up in lots of examples. Since a nonempty bipartite regular graph is degreeconstrained by definition, Theorem 12.5.6 guarantees it will have a perfect matching.
“mcs” — 2017/6/5 — 19:42 — page 474 — #482
474
Chapter 12
Simple Graphs
6:170
6:002
6:041
6:003
6:042
Figure 12.11 A scheduling graph for five exams. Exams connected by an edge
cannot be given at the same time.
12.6
Coloring
In Section 12.2, we used edges to indicate an affinity between a pair of nodes.
But there are lots of situations in which edges will correspond to conflicts between
nodes. Exam scheduling is a typical example.
12.6.1
An Exam Scheduling Problem
Each term, the MIT Schedules Office must assign a time slot for each final exam.
This is not easy, because some students are taking several classes with finals, and
(even at MIT) a student can take only one test during a particular time slot. The
Schedules Office wants to avoid all conflicts. Of course, you can make such a
schedule by having every exam in a different slot, but then you would need hundreds of slots for the hundreds of courses, and the exam period would run all year!
So, the Schedules Office would also like to keep exam period short.
The Schedules Office’s problem is easy to describe as a graph. There will be a
vertex for each course with a final exam, and two vertices will be adjacent exactly
when some student is taking both courses. For example, suppose we need to schedule exams for 6.041, 6.042, 6.002, 6.003 and 6.170. The scheduling graph might
appear as in Figure 12.11.
6.002 and 6.042 cannot have an exam at the same time since there are students in
both courses, so there is an edge between their nodes. On the other hand, 6.042 and
6.170 can have an exam at the same time if they’re taught at the same time (which
they sometimes are), since no student can be enrolled in both (that is, no student
should be enrolled in both when they have a timing conflict).
We next identify each time slot with a color. For example, Monday morning
“mcs” — 2017/6/5 — 19:42 — page 475 — #483
12.6. Coloring
475
blue
red
green
green
Figure 12.12
blue
A 3-coloring of the exam graph from Figure 12.11.
is red, Monday afternoon is blue, Tuesday morning is green, etc. Assigning an
exam to a time slot is then equivalent to coloring the corresponding vertex. The
main constraint is that adjacent vertices must get different colors—otherwise, some
student has two exams at the same time. Furthermore, in order to keep the exam
period short, we should try to color all the vertices using as few different colors as
possible. As shown in Figure 12.12, three colors suffice for our example.
The coloring in Figure 12.12 corresponds to giving one final on Monday morning
(red), two Monday afternoon (blue), and two Tuesday morning (green). Can we use
fewer than three colors? No! We can’t use only two colors since there is a triangle
in the graph, and three vertices in a triangle must all have different colors.
This is an example of a graph coloring problem: given a graph G, assign colors
to each node such that adjacent nodes have different colors. A color assignment
with this property is called a valid coloring of the graph—a “coloring,” for short.
A graph G is k-colorable if it has a coloring that uses at most k colors.
Definition 12.6.1. The minimum number of colors for which a graph G has a valid
coloring is called its chromatic number, .G/.
So G is k-colorable iff .G/ k.
12.6.2
Some Coloring Bounds
There are some simple properties of graphs that give useful bounds on colorability.
The simplest property is being a cycle: an even-length cycle is 2-colorable. So
.Ceven / D 2:
On the other hand, an odd-length cycle requires 3 colors, that is,
.Codd / D 3:
(12.4)
“mcs” — 2017/6/5 — 19:42 — page 476 — #484
476
Chapter 12
Simple Graphs
You should take a moment to think about why this equality holds.
Another simple example is a complete graph Kn :
.Kn / D n
since no two vertices can have the same color.
Being bipartite is another property closely related to colorability. If a graph is
bipartite, then you can color it with 2 colors using one color for the nodes on the
“left” and a second color for the nodes on the “right.” Conversely, graphs with
chromatic number 2 are all bipartite with all the vertices of one color on the “left”
and those with the other color on the right. Since only graphs with no edges—the
empty graphs—have chromatic number 1, we have:
Lemma 12.6.2. A nonempty graph G is bipartite iff .G/ D 2.
The chromatic number of a graph can also be shown to be small if the vertex
degrees of the graph are small. In particular, if we have an upper bound on the
degrees of all the vertices in a graph, then we can easily find a coloring with only
one more color than the degree bound.
Theorem 12.6.3. A graph with maximum degree at most n is .n C 1/-colorable.
Now you might be tempted to try to prove this theorem using induction on n,
but this would be a poor choice—we actually don’t know any way to this work,
so we expect it would ruin your week if this was your approach on a problem set.
On the other hand, the number of vertices or the number of edges are typically
good quantities on which to base induction proofs about graphs. We will prove
Theorem 12.6.3 by induction on the number of vertices.
Proof of Theorem 12.6.3. Let n 2 N be fixed. The induction hypothesis will be:
P .k/ WWD If a k-vertex graph has maximum degree n, then it is .n C 1/-colorable:
.
Base case (k D 1): A 1-vertex graph has maximum degree 0 and is 1-colorable, so
P .1/ is true.
Inductive step: Now assume that P .k/ is true, and let G be an .k C 1/-vertex
graph with maximum degree at most n. Remove a vertex v and all edges incident
to v from G, leaving a k-vertex subgraph H . The maximum degree of H is still at
most n, and so we can find an .n C 1/-coloring for H by induction hypothesis.
Now add back vertex v. Since there are more colors—n C 1—than vertices
adjacent to v—at most deg.v/ n—we can now pick a color for v that is different
from the colors of all these adjacent vertices. This yields a coloring of G with
.n C 1/-colors. So we have shown that G is .n C 1/-colorable, as required.
“mcs” — 2017/6/5 — 19:42 — page 477 — #485
12.6. Coloring
477
Figure 12.13 A seven-vertex star graph.
Sometimes nC1 colors is the best you can do. For example, every node in KnC1
has degree n and .KnC1 / D n C 1, so this is an example where Theorem 12.6.3
gives the best possible bound. Theorem 12.6.3 also gives the best possible coloring
bound for any graph with degree bounded by n that has KnC1 as a subgraph.
But sometimes n C 1 colors are far from the best that you can do. For example,
the star graph shown in Figure 12.13 has maximum degree six but can be colored
using just two colors.
In general it is very hard to find a minimal coloring for an arbitrary graph. In
fact, just calculating the chromatic number is hard. We’ll say more about this in the
next section.
12.6.3
Why coloring?
One reason coloring problems frequently arise in practice is because scheduling
conflicts are so common. For example, at the internet company Akamai, cofounded
by Tom Leighton, a new version of software is deployed over each of its servers
(200,000 servers in 2016) every few days. It would take more than twenty years to
update all these the servers one at a time, so the deployment must be carried out
for many servers simultaneouly. On the other hand, certain pairs of servers with
common critical functions cannot be updated simultaneouly, since a server needs
to be taken offline while being updated.
This problem gets solved by making a 200,000-node conflict graph and coloring
it with with a dozen or so colors—so only a dozen or so waves of installs are
needed!
Another example comes from the need to assign frequencies to radio stations. If
two stations have an overlap in their broadcast area, they can’t be given the same
frequency. Frequencies are precious and expensive, so it is important to minimize
the number handed out. This amounts to finding the minimum coloring for a graph
whose vertices are the stations and whose edges connect stations with overlapping
areas.
“mcs” — 2017/6/5 — 19:42 — page 478 — #486
478
Chapter 12
Simple Graphs
Coloring also comes up in allocating registers for program variables. While a
variable is in use, its value needs to be saved in a register. Registers can be reused
for different variables, but two variables need different registers if they are referenced during overlapping intervals of program execution. So register allocation is
the coloring problem for a graph whose vertices are the variables: vertices are adjacent if their intervals overlap, and the colors are registers. Once again, the goal is
to minimize the number of colors needed to color the graph.
Finally, there’s the famous map coloring problem stated in Proposition 1.1.4.
How many colors are needed to color a map so that adjacent territories get different
colors? Answering the question has no direct practical applications, but it tantalized graph theorists for over a century until an answer was discovered—to great
acclaim—in the 1970’s: four colors are enough to color any map. A reasonably
efficient way to find an actual 4-coloring can be based on that proof.
Surprisingly, the question whether there is an efficient procedure to tell if an
arbitrary planar graph really needs four colors—or if three would do the job—
turns out to be the SAT problem of Section 3.5 in disguise. So you win a million
dollars if you figure out how hard it is to determine 3-colorability. A proof that
testing 3-colorability of planar graphs is as hard as SAT is given in Problems 12.30
and 12.31.
12.7
Walks in Simple Graphs
12.7.1
Walks, Paths, Cycles
Walks and paths in simple graphs are esentially the same as in digraphs. We just
modify the digraph definitions using undirected edges instead of directed ones. For
example, the formal definition of a walk in a simple graph is a virtually the same
as the Definition 10.2.1 of a walk in a digraph:
Definition 12.7.1. A walk in a simple graph G is an alternating sequence of vertices
and edges that begins with a vertex, ends with a vertex, and such that for every edge
hu—vi in the walk, one of the endpoints u, v is the element just before the edge,
and the other endpoint is the next element after the edge. The length of a walk is
the total number of occurrences of edges in it.
So a walk v is a sequence of the form
v WWD v0 hv0 —v1 i v1 hv1 —v2 i v2 : : : hvk
1 —vk i
vk
where hvi —vi C1 i 2 E.G/ for i 2 Œ0::k/. The walk is said to start at v0 , to end
“mcs” — 2017/6/5 — 19:42 — page 479 — #487
12.7. Walks in Simple Graphs
479
d
b
a
e
c
g
h
f
Figure 12.14 A graph with 3 cycles: bhecb, cdec, bcdehb.
at vk , and the length, jvj, of the walk is k. The walk is a path iff all the vi ’s are
different, that is, if i ¤ j , then vi ¤ vj .
A walk that begins and ends at the same vertex is a closed walk. A single vertex
counts as a length zero closed walk as well as a length zero path.
A cycle can be represented by a closed walk of length three or more whose
vertices are distinct except for the beginning and end vertices.
Note that in contrast to digraphs, we don’t count length two closed walks as
cycles in simple graphs. That’s because a walk going back and forth on the same
edge is neither interesting nor important. There are also no closed walks of length
one, since simple graphs don’t have self-loops.
As in digraphs, the length of a walk is the number of occurrences of edges in the
walk, which is one less than the number of occurrences of vertices. For example,
the graph in Figure 12.14 has a length 6 path through the seven successive vertices
abcdefg. This is the longest path in the graph. The graph in Figure 12.14 also has
three cycles through successive vertices bhecb, cdec and bcdehb.
12.7.2
Cycles as Subgraphs
We don’t intend that cycles have a beginning or an end, so any of the paths that go
around a cycle can represent it. For example, in the graph in Figure 12.14, the cycle
starting at b and going through vertices bcdehb can also be described as starting
at d and going through dehbcd . Furthermore, cycles in simple graphs don’t have
a direction: dcbhed describes the same cycle as though it started and ended at d
but went in the opposite direction.
A precise way to explain which closed walks represent the same cycle is to define
a cycle to be subgraph isomorphic to a cycle graph Cn .
“mcs” — 2017/6/5 — 19:42 — page 480 — #488
480
Chapter 12
Simple Graphs
Definition 12.7.2. A graph G is said to be a subgraph of a graph H if V .G/ V .H / and E.G/ E.H /.
For example, the one-edge graph G where
V .G/ D fg; h; i g
and
E.G/ D f hh—i i g
is a subgraph of the graph H in Figure 12.1. On the other hand, any graph containing an edge hg—hi will not be a subgraph of H because this edge is not in
E.H /. Another example is an empty graph on n nodes, which will be a subgraph
of an Ln with the same set of vertices; similarly, Ln is a subgraph of Cn , and Cn
is a subgraph of Kn .
Definition 12.7.3. For n 3, let Cn be the graph with vertices Œ0::n/ and edges
h0—1i ; h1—2i ; : : : ; h.n
2/—n
1i ; hn
1—0i :
A cycle of a graph G is a subgraph of G that is isomorphic to Cn for some n 3.
This definition formally captures the idea that cycles don’t have direction or beginnings or ends.
12.8
Connectivity
Definition 12.8.1. Two vertices are connected in a graph when there is a path that
begins at one and ends at the other. By convention, every vertex is connected to
itself by a path of length zero. A graph is connected when every pair of vertices
are connected.
12.8.1
Connected Components
Being connected is usually a good property for a graph to have. For example, it
could mean that it is possible to get from any node to any other node, or that it is
possible to communicate between any pair of nodes, depending on the application.
But not all graphs are connected. For example, the graph where nodes represent
cities and edges represent highways might be connected for North American cities,
but would surely not be connected if you also included cities in Australia. The
same is true for communication networks like the internet—in order to be protected
from viruses that spread on the internet, some government networks are completely
isolated from the internet.
“mcs” — 2017/6/5 — 19:42 — page 481 — #489
12.8. Connectivity
481
Figure 12.15
One graph with 3 connected components.
Another example is shown in Figure 12.15, which looks like a picture of three
graphs, but is intended to be a picture of one graph. This graph consists of three
pieces. Each piece is a subgraph that by itself is connected, but there are no paths
between vertices in different pieces. These connected pieces of a graph are called
its connected components.
Definition 12.8.2. A connected component of a graph is a subgraph consisting of
some vertex and every node and edge that is connected to that vertex.
So a graph is connected iff it has exactly one connected component. At the
other extreme, the empty graph on n vertices has n connected components, each
consisting of a single vertex.
12.8.2
Odd Cycles and 2-Colorability
We have already seen that determining the chromatic number of a graph is a challenging problem. There is one special case where this problem is very easy, namely,
when the graph is 2-colorable.
Theorem 12.8.3. The following graph properties are equivalent:
1. The graph contains an odd length cycle.
2. The graph is not 2-colorable.
3. The graph contains an odd length closed walk.
In other words, if a graph has any one of the three properties above, then it has
all of the properties.
We will show the following implications among these properties:
1. IMPLIES 2. IMPLIES 3. IMPLIES 1:
So each of these properties implies the other two, which means they all are equivalent.
“mcs” — 2017/6/5 — 19:42 — page 482 — #490
482
Chapter 12
Simple Graphs
1. IMPLIES 2. Proof. This follows from equation 12.4.
2. IMPLIES 3. If we prove this implication for connected graphs, then it will hold
for an arbitrary graph because it will hold for each connected component. So
we can assume that G is connected.
Proof. Pick an arbitrary vertex r of G. Since G is connected, for every node
u 2 V .G/, there will be a walk wu starting at u and ending at r. Assign
colors to vertices of G as follows:
(
black; if jwu j is even;
color.u/ D
white; otherwise:
Now since G is not colorable, this can’t be a valid coloring. So there must
be an edge between two nodes u and v with the same color. But in that case
wu breverse.wv /b hv—ui
is a closed walk starting and ending at u, and its length is
jwu j C jwv j C 1
which is odd.
3. IMPLIES 1. Proof. Since there is an odd length closed walk, the WOP implies
there is an odd length closed walk w of minimum length. We claim w must
be a cycle. To show this, assume to the contrary that w is not a cycle, so
there is a repeat vertex occurrence besides the start and end. There are then
two cases to consider depending on whether the additional repeat is different
from, or the same as, the start vertex.
In the first case, the start vertex has an extra occurrence. That is,
w D fb
xr
for some positive length walks f and r that begin and end at x. Since
jwj D jfj C jrj
is odd, exactly one of f and r must have odd length, and that one will be an
odd length closed walk shorter than w, a contradiction.
In the second case,
w D fb
y gb
yr
“mcs” — 2017/6/5 — 19:42 — page 483 — #491
12.9. Special Walks and Tours
Figure 12.16
483
Is there a walk that includes every edge exactly once?
where f is a walk from x to y for some y ¤ x, and r is a walk from y to
x, and jgj > 0. Now g cannot have odd length or it would be an odd-length
closed walk shorter than w. So g has even length. That implies that fb
y r must
be an odd-length closed walk shorter than w, again a contradiction.
This completes the proof of Theorem 12.8.3.
12.9
Special Walks and Tours
12.9.1
Euler Tours
Can you walk every hallway in the Museum of Fine Arts exactly once? If we represent hallways with vertices and connections between hallways with edges, then
this reduces to a question about graphs. For example, could you visit every hallway
exactly once in a museum with floor plan represented by the graph in Figure 12.16?
The entire field of graph theory began when the famous 17th century mathematician Leonhard Euler5 asked whether the seven bridges of Königsberg could all be
crossed exactly once—essentially the same question we asked about the Museum
of Fine Arts.
An Euler tour6 of a graph is a closed walk that includes every edge exactly once.
Graphs with Euler tours turn out to be easy to recognize:
5 Same
6 In
Euler who defined the constant e 2:718 and the totient function —he did a lot of stuff.
some other texts, Euler “tours” are called Euler “circuits.”
“mcs” — 2017/6/5 — 19:42 — page 484 — #492
484
Chapter 12
Simple Graphs
Theorem 12.9.1. A connected graph has an Euler tour if and only if every vertex
has even degree.
The proof is worked out in a series of easy steps in Problem 12.43. Implicit in
the proof is a simple way to find Euler tours when they exist, namely, grow a tour
by repeatedly splicing in closed walks until all the edges are included.
Similarly, there is a walk that includes every edge exactly once but is not a tour—
it begins and ends at different verticies—if and only if there are exactly two vertices
of odd degree. The graph shown in Figure 12.16 has exactly two vertices with odd
degree, so it has a such a walk but not an Euler tour.
12.9.2
Hamiltonian Cycles
Hamiltonian cycles are the unruly cousins of Euler tours.
Definition 12.9.2. A Hamiltonian cycle in a graph G is a cycle that visits every
vertex in G exactly once. Similarly, a Hamiltonian path is a path in G that visits
every vertex exactly once.
Although Hamiltonian cycles sound similar to Euler tours—one visits every node
once while the other visits every edge once—finding a Hamiltonian cycle or a
Hamiltonian path can be a lot harder than finding an Euler tour or path. No one
has yet discovered a simple characterization for when a graph has a Hamiltonian
cycle. In fact, determining whether a graph has a Hamiltonian cycle is the same category of problem as the SAT problem of Section 3.5: you get a million dollars from
the Clay Institute for finding an efficient way to determine whether a graph has a
Hamiltonian cycle—or for proving that there is no procedure that works efficiently
on all graphs.
12.9.3
The Traveling Salesperson Problem
In many applications, there are numerical costs or weights associated with the edges
of the graph. For example, suppose the nodes of a graph represent buildings and
edges represent connections between them. The cost of a connection may vary a
lot from one pair of buildings or towns to another. Another example is where the
nodes represent cities, and the weight of an edge is the distance between them: the
weight of the Los Angeles/New York City edge is much higher than the weight of
the NYC/Boston edge.
As if the problem of finding a Hamiltonian cycle is not hard enough, in a weighted
graph, we sometimes need to find a Hamiltonian cycle with minimum possible sum
of the weights of its edges. This is the famous Traveling Salesperson Problem.
“mcs” — 2017/6/5 — 19:42 — page 485 — #493
12.10. k-connected Graphs
485
3
v2
4
v3
3
1
2
v1
3
4
5
2
v6
Figure 12.17
v4
6
v5
Is there a Hamiltonian cycle with weight 15?
For example, suppose you would like to visit every node in the graph shown in
Figure 12.17 exactly once, finishing at the node where you started. Can you find a
weight 15 cycle that does this?
12.10 k-connected Graphs
If we think of a graph as modeling cables in a telephone network, or oil pipelines,
or electrical power lines, then we not only want connectivity, but we want connectivity that survives component failure. One measure of connection strength is how
many links must fail before connectedness fails. In particular, two vertices are kedge connected when it takes at least k “edge-failures” to disconnect them. More
precisely:
Definition 12.10.1. Two vertices in a graph are k-edge connected when they remain connected in every subgraph obtained by deleting up to k 1 edges. A graph
with two or more vertices is k-edge connected when every pair of distinct vertices
in the graph are k-edge connected.
From now on we’ll drop the “edge” modifier and just say “k-connected.”7
For example, in the graph in figure 12.14, vertices c and e are 3-connected, b
and e are 2-connected, g and e are 1 connected, and no vertices are 4-connected.
7 There
is a corresponding definition of k-vertex connectedness based on deleting vertices rather
than edges. Graph theory texts usually use “k-connected” as shorthand for “k-vertex connected.” But
edge-connectedness will be enough for us.
“mcs” — 2017/6/5 — 19:42 — page 486 — #494
486
Chapter 12
Simple Graphs
The graph as a whole is only 1-connected. A complete graph Kn is .n 1/connected. Every cycle is 2-connected.
Notice that two vertices are 1-connected iff they are connected according to Definition 12.8.1.
The idea of a cut edge is a useful way to explain 2-connectivity.
Definition 12.10.2. If two vertices are connected in a graph G, but not connected
when an edge e is removed, then e is called a cut edge of G.
So a graph with more than one vertex is 2-connected iff it is connected and
has no cut edges. The following Lemma is another immediate consequence of the
definition:
Lemma 12.10.3. An edge is a cut edge iff it is not on a cycle.
More generally, if two vertices are connected by k edge-disjoint paths—that is,
no edge occurs in two paths—then they must be k-connected, since at least one
edge will have to be removed from each of the paths before they could disconnect.
A fundamental fact, whose ingenious proof we omit, is Menger’s theorem which
confirms that the converse is also true: if two vertices are k-connected, then there
are k edge-disjoint paths connecting them. It takes some ingenuity to prove this
just for the case k D 2.
12.10.1
The Minimum Number of Edges in a Connected Graph
The following theorem says that a graph with few edges must have many connected
components.
Theorem 12.10.4. Every graph G has at least jV .G/j
ponents.
jE.G/j connected com-
Of course for Theorem 12.10.4 to be of any use, there must be fewer edges than
vertices.
Proof. We use induction on the number k of edges. Let P .k/ be the proposition
that
every graph G with k edges has at least jV .G/j
nents.
k connected compo-
Base case (k D 0): In a graph with 0 edges, each vertex is itself a connected
component, and so there are exactly jV .G/j D jV .G/j 0 connected components.
So P .0/ holds.
“mcs” — 2017/6/5 — 19:42 — page 487 — #495
12.11. Forests & Trees
487
Inductive step:
Let Ge be the graph that results from removing an edge, e 2 E.G/. So Ge
has k edges, and by the induction hypothesis P .k/, we may assume that Ge has
at least .jV .G/j k/ connected components. Now add back the edge e to obtain
the original graph G. If the endpoints of e were in the same connected component
of Ge , then G has the same sets of connected vertices as Ge , so G has at least
.jV .G/j k/ > .jV .G/j .k C 1// components. Alternatively, if the endpoints
of e were in different connected components of Ge , then these two components are
merged into one component in G, while all other components remain unchanged,
so that G has one fewer connected component than Ge . That is, G has at least
.jV .G/j k/ 1 D .jV .G/j .k C 1// connected components. So in either case,
G has at least jV .G/j .k C 1/ components, as claimed.
This completes the inductive step and hence the entire proof by induction.
Corollary 12.10.5. Every connected graph with n vertices has at least n 1 edges.
A couple of points about the proof of Theorem 12.10.4 are worth noticing. First,
we used induction on the number of edges in the graph. This is very common in
proofs involving graphs, as is induction on the number of vertices. When you’re
presented with a graph problem, these two approaches should be among the first
you consider.
The second point is more subtle. Notice that in the inductive step, we took an
arbitrary .k C1/-edge graph, threw out an edge so that we could apply the induction
assumption, and then put the edge back. You’ll see this shrink-down, grow-back
process very often in the inductive steps of proofs related to graphs. This might
seem like needless effort: why not start with an k-edge graph and add one more to
get an .k C 1/-edge graph? That would work fine in this case, but opens the door
to a nasty logical error called buildup error, illustrated in Problem 12.40.
12.11
Forests & Trees
We’ve already made good use of acyclic digraphs, but acyclic simple graphs are
arguably the most important graphs in computer science.
Definition 12.11.1. An acyclic graph is called a forest. A connected acyclic graph
is called a tree.
The graph shown in Figure 12.18 is a forest. Each of its connected components
is by definition a tree.
“mcs” — 2017/6/5 — 19:42 — page 488 — #496
488
Chapter 12
Simple Graphs
Figure 12.18 A 6-node forest consisting of 2 component trees.
a
e
h
c
b
g
d
i
f
Figure 12.19
12.11.1
A 9-node tree with 5 leaves.
Leaves, Parents & Children
One of the first things you will notice about trees is that they tend to have a lot of
nodes with degree one.
Definition 12.11.2. A degree 1 node in a forest is called a leaf.
The forest in Figure 12.18 has 4 leaves. The tree in Figure 12.19 has 5 leaves.
12.11.2
Tree Properties
Trees have many unique properties. We have listed some of them in the following
theorem.
Theorem 12.11.3. Every tree has the following properties:
1. Every connected subgraph is a tree.
2. There is a unique path between every pair of vertices.
3. Adding an edge between any two nonadjacent nodes in a tree creates a graph
with a cycle.
4. Removing any edge disconnects the graph. That is, every edge is a cut edge.
5. If the tree has at least two vertices, then it has at least two leaves.
6. The number of vertices in a tree is one larger than the number of edges.
“mcs” — 2017/6/5 — 19:42 — page 489 — #497
12.11. Forests & Trees
Proof.
489
1. A cycle in a subgraph is also a cycle in the whole graph, so any subgraph of an acyclic graph must also be acyclic. If the subgraph is also connected, then by definition, it is a tree.
2. Since a tree is connected, there is at least one path between every pair of vertices. Suppose for the purposes of contradiction, that there are two different
paths between some pair of vertices. Then there are two distinct paths p ¤ q
between the same two vertices with minimum total length jpj C jqj. If these
paths shared a vertex w other than at the start and end of the paths, then the
parts of p and q from start to w, or the parts of p and q from w to the end,
must be distinct paths between the same vertices with total length less than
jpj C jqj, contradicting the minimality of this sum. Therefore, p and q have
no vertices in common besides their endpoints, and so p breverse.q/ is a
cycle.
3. An additional edge hu—vi together with the unique path between u and v
forms a cycle.
4. Suppose that we remove edge hu—vi. Since the tree contained a unique path
between u and v, that path must have been hu—vi. Therefore, when that
edge is removed, no path remains, and so the graph is not connected.
5. Since the tree has at least two vertices, the longest path in the tree will have
different endpoints u and v. We claim u is a leaf. This follows because, by
definition of endpoint, u is incident to at most one edge on the path. Also, if u
was incident to an edge not on the path, then the path could be lengthened by
adding that edge, contradicting the fact that the path was as long as possible.
It follows that u is incident only to a single edge, that is u is a leaf. The same
hold for v.
6. We use induction on the proposition
P .n/ WWD there are n
1 edges in any n-vertex tree:
Base case (n D 1): P .1/ is true since a tree with 1 node has 0 edges and
1 1 D 0.
Inductive step: Now suppose that P .n/ is true and consider an .nC1/-vertex
tree T . Let v be a leaf of the tree. You can verify that deleting a vertex of
degree 1 (and its incident edge) from any connected graph leaves a connected
subgraph. So by Theorem 12.11.3.1, deleting v and its incident edge gives
“mcs” — 2017/6/5 — 19:42 — page 490 — #498
490
Chapter 12
Simple Graphs
Figure 12.20 A graph where the edges of a spanning tree have been thickened.
a smaller tree, and this smaller tree has n 1 edges by induction. If we reattach the vertex v and its incident edge, we find that T has n D .n C 1/ 1
edges. Hence, P .n C 1/ is true, and the induction proof is complete.
Various subsets of properties in Theorem 12.11.3 provide alternative characterizations of trees. For example,
Lemma 12.11.4. A graph G is a tree iff G is a forest and jV .G/j D jE.G/j C 1.
The proof is an easy consequence of Theorem 12.11.3.6, see Problem 12.48.
12.11.3
Spanning Trees
Trees are everywhere. In fact, every connected graph contains a subgraph that is a
tree with the same vertices as the graph. This is called a spanning tree for the graph.
For example, Figure 12.20 is a connected graph with a spanning tree highlighted.
Definition 12.11.5. A spanning subgraph of a graph G is a subgraph whose vertices are the same as the vertices of G.
Theorem 12.11.6. Every connected graph contains a spanning tree.
Proof. Suppose G is a connected graph, so the graph G itself is a connected, spanning subgraph. So by WOP, G must have a minimum-edge connected, spanning
subgraph T .
We claim T is a spanning tree. Since T is a connected, spanning subgraph by
definition, all we have to show is that T is acyclic. But suppose to the contrary that
T contained a cycle C . By Lemma 12.10.3, an edge e of C will not be a cut edge,
so removing it would leave a connected, spanning subgraph that was smaller than
T , contradicting the minimality to T .
“mcs” — 2017/6/5 — 19:42 — page 491 — #499
12.11. Forests & Trees
491
1
2
2
1
3
2
3
2
3
1
3
4
1
1
3
1
7
7
(a)
(b)
Figure 12.21 A spanning tree (a) with weight 19 for a graph (b).
1
2
2
1
1
3
7
Figure 12.22
12.11.4
An MST with weight 17 for the graph in Figure 12.21(b).
Minimum-Weight Spanning Trees
Spanning trees are particularly interesting and important for weighted graphs—
graphs whose edges are assigned numerical weights. The overall weight of a
weighted graph is the sum of the weights of all its edges. For example, the weight
of the spanning tree shown in Figure 12.21(a) is 19.
The general goal is to find a spanning tree with minimum weight, called a
minimum-weight spanning tree (MST).
The spanning tree shown in Figure 12.21(a) turns out not to be an MST of the
graph in Figure 12.21(b): Figure 12.22 shows another spanning tree for this graph
with weight 17. This weight 17 turns out to be an MST, but how do we prove it? In
general, how do we find an MST for a connected graph G? We could try enumerating all subtrees of G, but that approach would be hopeless for large graphs.
For the rest of this section, suppose G is a connected weighted graph. We’ll now
explain a simple, general way to find an MST of G.
“mcs” — 2017/6/5 — 19:42 — page 492 — #500
492
Chapter 12
Simple Graphs
Definition 12.11.7. A black-white coloring of a simple graph G is a partition of
the vertices V .G/ into two blocks—the “black” vertices and the “white” vertices.
A gray edge of the coloring is an edge of G with different colored endpoints.
Lemma 12.11.8. If G is a connected graph, then every black-white coloring of G
contains a gray edge.
Proof. Any coloring of G must have at least one vertex of each color—this follows
because the blocks of a partition are by definition nonempty.
Since G is connected, there will be a path that begins at a black vertex and ends
at a white vertex. Therefore at some point on the path there has to be an edge that
starts with a black vertex and ends with a white one. This is the gray edge.
The weighted graphs examples up to now each had two or more edges with the
same weight. This is common. But our story becomes simpler if all weights are
different.
For the rest of this section, assume there are no same-weight edges:
different edges have different weights.
There is really no loss of generality in assuming different weights, since everything we do below assuming distinct weights would work if we assigned an arbitrary rank to same-weight edges and treated higher rank vertices as having larger
weight.
The following lemma now makes it easy to find MST’s.
Lemma 12.11.9 (Gray Edge Lemma). For any black-white coloring of a connected
graph G, the minimum-weight edge among the gray edges of the coloring must be
an edge of every MST of G.
So to find an MST for G, we just need to find enough minimum-weight gray
edges to form a spnning tree. We do this using:
Minimum-Weight Gray Edge Procedure
Start off with the disconnected spanning forest consisting of the vertices of G with all edges removed. Next, suppose S is a spanning forest
built up at some point by the procedure. If S is not yet connected, then
pick a black-white coloring of G such that such that all the vertices in
the same connected component of S have the same color. Update S
by adding the minimum weight gray edge of this coloring to the edges
of S . Continue until S becomes connected.
“mcs” — 2017/6/5 — 19:42 — page 493 — #501
12.11. Forests & Trees
493
Theorem 12.11.10. The Minimum-Weight Gray Edge Procedure will terminate
with an MST.
Proof. To see why this procedure works, suppose G has n vertices and let S be
the current spanning forest at any point in the procedure. If S is not connected, a
black-white coloring is chosen so that the endpoints of any edge of S are the same
color. This ensures that no gray edge is an edge of S .
By Lemma 12.11.8, there must be at least one gray edge in the coloring, so
there will be a minimum-weight gray edge.8 Updating S by adding this gray edge
decreases the number of connected components of S by one, so after adding n 1
gray edges, the procedure will terminate with a connected spanning forest, that is,
a spanning tree T .
By the Gray Edge Lemma 12.11.9, every edge of T has to be in every MST of
G. This is only possible if all MST’s have the same edges as T , that is, they all
equal T . So T is the unique possible MST of G.
Corollary 12.11.11. Every connected graph with no same-weight edges has a
unique MST.
Here are three known algorithms that are special cases of the Minimum Weight
Gray Edge Procedure;
Algorithm 1. [Prim] Grow S one edge at a time by adding a minimum weight
edge among the edges that have exactly one endpoint in S .
This is the algorithm that comes from coloring the growing tree white and all the
vertices not in the tree black. Then the gray edges are the ones with exactly one
endpoint in the tree.
Algorithm 2. [Kruskal] Grow S one edge at a time by adding a minimum weight
edge among the edges that do not create a cycle in S .
An edge does not create a cycle in S iff it connects different components of S .
So the edge chosen by Kruskal’s algorithm will be the minimum weight gray edge
when the components it connects are assigned different colors.
The coloring that explains Algorithm 1 also justifies a more flexible algorithm
which has Algorithm 1 as a special case:
Algorithm 3. Grow a forest one edge at a time by picking any component and
adding a minimum weight edge among the edges leaving that component.
8 . . . by
the Well Ordering Principle for finite sets Lemma2.4.5.
“mcs” — 2017/6/5 — 19:42 — page 494 — #502
494
Chapter 12
Simple Graphs
This algorithm allows components that are not too close to grow in parallel and
independently, which is great for “distributed” computation where separate processors share the work with limited communication between processors.9
Ok, to wrap up this story, all that’s left is the proof of Lemma 12.11.9 that minweight gray edges must be in every spanning tree. Let’s start with another lemma:
Lemma 12.11.12. [Gray Edge Swap] Let G be a connected weighted graph, and
let e be the minimum-weight gray edge of some black-white coloring of of G. Suppose C is a connected spanning subgraph of G, and e is not an edge of C . Then
there is an edge g of C such that
weight.e/ < weight.g/,
C g C e is connected, where C g C e is the graph obtained by removing
g and adding e to the set of edges of C .
Notice that the Gray Edge Lemma 12.11.9 now follows immediately from the
Swap Lemma 12.11.12: if C is a spanning tree of G, then C g C e will also be a
spanning tree, and it has smaller weight. So C cannot be minimum-weight unless
e is one of its edges.
Proof. (Swap Lemma 12.11.12) Pick a black-white coloring for which e is a minweight gray edge. Since C is a connected spanning graph, there is a path p in C
between the different colored endpoints of e. Now p has both a black endpoint
and a white one, so it must contain some gray edge g ¤ e. Since e is a minimum
weight gray edge and there are no same-weight edges, we have
weight.e/ < weight.g/:
Next, notice that p C e is a cycle. So when g is removed from the cycle, the
endpoints of g remain connected by the rest of the cycle. This means C C e g is
still a connected, spanning subgraph, as required.
12.12
References
[9], [15], [24], [27], [29]
9 The idea of growing trees in parallel seems first to have been developed in by Boru̇vka (1926), ref
TBA. Efficient MST algorithms running in parallel time O.log jV j/ are described in Karger, Klein,
and Tarjan (1995), ref TBA.
“mcs” — 2017/6/5 — 19:42 — page 495 — #503
12.12. References
495
Problems for Section 12.2
Practice Problems
Problem 12.1.
The average degree of the vertices in an n-vertex graph is twice the average number
of edges per vertex. Explain why.
Problem 12.2.
Among connected simple graphs whose sum of vertex degrees is 20:
(a) what is the largest possible number of vertices?
(b) what is the smallest possible number of vertices?
Class Problems
Problem 12.3. (a) Prove that in every simple graph, there are an even number of
vertices of odd degree.
(b) Conclude that at a party where some people shake hands, the number of people
who shake hands an odd number of times is an even number.
(c) Call a sequence of people at the party a handshake sequence if each person in
the sequence has shaken hands with the next person, if any, in the sequence.
Suppose George was at the party and has shaken hands with an odd number of
people. Explain why, starting with George, there must be a handshake sequence
ending with a different person who has shaken an odd number of hands.
Exam Problems
Problem 12.4.
A researcher analyzing data on heterosexual sexual behavior in a group of m males
and f females found that within the group, the male average number of female
partners was 10% larger that the female average number of male partners.
(a) Comment on the following claim. “Since we’re assuming that each encounter
involves one man and one woman, the average numbers should be the same, so the
males must be exaggerating.”
(b) For what constant c is m D c f ?
“mcs” — 2017/6/5 — 19:42 — page 496 — #504
496
Chapter 12
Simple Graphs
(c) The data shows that approximately 20% of the females were virgins, while
only 5% of the males were. The researcher wonders how excluding virgins from
the population would change the averages. If he knew graph theory, the researcher
would realize that the nonvirgin male average number of partners will be x.f =m/
times the nonvirgin female average number of partners. What is x?
(d) For purposes of further research, it would be helpful to pair each female in the
group with a unique male in the group. Explain why this is not possible.
Problems for Section 12.4
Practice Problems
Problem 12.5.
Which of the items below are simple-graph properties preserved under isomorphism?
(a) There is a cycle that includes all the vertices.
(b) The vertices are numbered 1 through 7.
(c) The vertices can be numbered 1 through 7.
(d) There are two degree 8 vertices.
(e) Two edges are of equal length.
(f) No matter which edge is removed, there is a path between any two vertices.
(g) There are two cycles that do not share any vertices.
(h) The vertices are sets.
(i) The graph can be drawn in a way that all the edges have the same length.
(j) No two edges cross.
(k) The OR of two properties that are preserved under isomorphism.
(l) The negation of a property that is preserved under isomorphism.
“mcs” — 2017/6/5 — 19:42 — page 497 — #505
12.12. References
497
Class Problems
Problem 12.6.
For each of the following pairs of simple graphs, either define an isomorphism
between them, or prove that there is none. (We write ab as shorthand for ha—bi.)
(a)
G1 with V1 D f1; 2; 3; 4; 5; 6g; E1 D f12; 23; 34; 14; 15; 35; 45g
G2 with V2 D f1; 2; 3; 4; 5; 6g; E2 D f12; 23; 34; 45; 51; 24; 25g
(b)
G3 with V3 D f1; 2; 3; 4; 5; 6g; E3 D f12; 23; 34; 14; 45; 56; 26g
G4 with V4 D fa; b; c; d; e; f g; E4 D fab; bc; cd; de; ae; ef; cf g
Problem 12.7.
List all the isomorphisms between the two graphs given in Figure 12.23. Explain
why there are no others.
1
a
2
3
5
4
6
b
c
e
d
f
Figure 12.23 Graphs with several isomorphisms
Homework Problems
Problem 12.8.
Determine which among the four graphs pictured in Figure 12.24 are isomorphic.
For each pair of isomorphic graphs, describe an isomorphism between them. For
each pair of graphs that are not isomorphic, give a property that is preserved under
isomorphism such that one graph has the property, but the other does not. For
at least one of the properties you choose, prove that it is indeed preserved under
isomorphism (you only need prove one of them).
“mcs” — 2017/6/5 — 19:42 — page 498 — #506
498
Chapter 12
Simple Graphs
1
1
6
5
6
8
2
9
5
9
7
10
8
10
3
4
2
7
3
4
(a) G1
(b) G2
1
1
9
2
6
5
9
2
7
8
10
(c) G3
3
10
7
3
4
8
4
5
6
(d) G4
Figure 12.24 Which graphs are isomorphic?
“mcs” — 2017/6/5 — 19:42 — page 499 — #507
12.12. References
499
Problem 12.9. (a) For any vertex v in a graph, let N.v/ be the set of neighbors of
v, namely, the vertices adjacent to v:
N.v/ WWD fu j hu—vi is an edge of the graphg:
Suppose f is an isomorphism from graph G to graph H . Prove that f .N.v// D
N.f .v//.
Your proof should follow by simple reasoning using the definitions of isomorphism
and neighbors—no pictures or handwaving.
Hint: Prove by a chain of iff’s that
h 2 N.f .v// iff
h 2 f .N.v//
for every h 2 VH . Use the fact that h D f .u/ for some u 2 VG .
(b) Conclude that if G and H are isomorphic graphs, then for each k 2 N, they
have the same number of degree k vertices.
Problem 12.10.
Let’s say that a graph has “two ends” if it has exactly two vertices of degree 1 and
all its other vertices have degree 2. For example, here is one such graph:
(a) A line graph is a graph whose vertices can be listed in a sequence with edges
between consecutive vertices only. So the two-ended graph above is also a line
graph of length 4.
Prove that the following theorem is false by drawing a counterexample.
False Theorem. Every two-ended graph is a line graph.
(b) Point out the first erroneous statement in the following bogus proof of the false
theorem and describe the error.
Bogus proof. We use induction. The induction hypothesis is that every two-ended
graph with n edges is a line graph.
Base case (n D 1): The only two-ended graph with a single edge consists of two
vertices joined by an edge:
“mcs” — 2017/6/5 — 19:42 — page 500 — #508
500
Chapter 12
Simple Graphs
Sure enough, this is a line graph.
Inductive case: We assume that the induction hypothesis holds for some n 1
and prove that it holds for n C 1. Let Gn be any two-ended graph with n edges.
By the induction assumption, Gn is a line graph. Now suppose that we create a
two-ended graph GnC1 by adding one more edge to Gn . This can be done in only
one way: the new edge must join one of the two endpoints of Gn to a new vertex;
otherwise, GnC1 would not be two-ended.
gn
↑
new edge
Clearly, GnC1 is also a line graph. Therefore, the induction hypothesis holds for
all graphs with n C 1 edges, which completes the proof by induction.
Problems for Section 12.5
Practice Problems
Problem 12.11.
Let B be a bipartite graph with vertex sets L.B/; R.B/. Explain why the sum of
the degrees of the vertices in L.B/ equals the sum of the degrees of the vertices in
R.B/.
Class Problems
Problem 12.12.
A certain Institute of Technology has a lot of student clubs; these are loosely overseen by the Student Association. Each eligible club would like to delegate one of its
members to appeal to the Dean for funding, but the Dean will not allow a student to
be the delegate of more than one club. Fortunately, the Association VP took Math
for Computer Science and recognizes a matching problem when she sees one.
“mcs” — 2017/6/5 — 19:42 — page 501 — #509
12.12. References
501
(a) Explain how to model the delegate selection problem as a bipartite matching
problem. (This is a modeling problem; we aren’t looking for a description of an
algorithm to solve the problem.)
(b) The VP’s records show that no student is a member of more than 9 clubs. The
VP also knows that to be eligible for support from the Dean’s office, a club must
have at least 13 members. That’s enough for her to guarantee there is a proper
delegate selection. Explain. (If only the VP had taken an Algorithms class, she
could even have found a delegate selection without much effort.)
Problem 12.13.
A simple graph is called regular when every vertex has the same degree. Call
a graph balanced when it is regular and is also a bipartite graph with the same
number of left and right vertices.
Prove that if G is a balanced graph, then the edges of G can be partitioned into
blocks such that each block is a perfect matching.
For example, if G is a balanced graph with 2k vertices each of degree j , then the
edges of G can be partitioned into j blocks, where each block consists of k edges,
each of which is a perfect matching.
Exam Problems
Problem 12.14.
Overworked and over-caffeinated, the Teaching Assistant’s (TA’s) decide to oust
the lecturer and teach their own recitations. They will run a recitation session at 4
different times in the same room. There are exactly 20 chairs to which a student can
be assigned in each recitation. Each student has provided the TA’s with a list of the
recitation sessions her schedule allows and each student’s schedule conflicts with
at most two sessions. The TA’s must assign each student to a chair during recitation
at a time she can attend, if such an assignment is possible.
(a) Describe how to model this situation as a matching problem. Be sure to specify what the vertices/edges should be and briefly describe how a matching would
determine seat assignments for each student in a recitation that does not conflict
with his schedule. (This is a modeling problem; we aren’t looking for a description
of an algorithm to solve the problem.)
(b) Suppose there are 41 students. Given the information provided above, is a
matching guaranteed? Briefly explain.
“mcs” — 2017/6/5 — 19:42 — page 502 — #510
502
Chapter 12
Simple Graphs
Problem 12.15.
Because of the incredible popularity of his class Math for Computer Science, TA
Mike decides to give up on regular office hours. Instead, he arranges for each
student to join some study groups. Each group must choose a representative to talk
to the staff, but there is a staff rule that a student can only represent one group. The
problem is to find a representative from each group while obeying the staff rule.
(a) Explain how to model the delegate selection problem as a bipartite matching
problem. (This is a modeling problem; we aren’t looking for a description of an
algorithm to solve the problem.)
(b) The staff’s records show that each student is a member of at most 4 groups,
and all the groups have 4 or more members. Is that enough to guarantee there is a
proper delegate selection? Explain.
Problem 12.16.
b be the “implies” binary relation on propositional formulas defined by the rule
Let R
that
b G iff Œ.F IMPLIES G/ is a valid formula:
F R
(12.5)
b P , because the formula .P AND Q/ IMPLIES P is
For example, .P AND Q/ R
b P since .P OR Q/ IMPLIES P is not
valid. Also, it is not true that .P OR Q/ R
valid.
b is not a weak
(a) Let A and B be the sets of formulas listed below. Explain why R
partial order on the set A [ B.
b arrows from A to B.
(b) Fill in the R
“mcs” — 2017/6/5 — 19:42 — page 503 — #511
12.12. References
503
A
B
Q
arrows
P XOR Q
P OR Q
P AND Q
P OR Q OR .P AND Q/
NOT .P AND
Q/
P
(c) The diagram in part (b) defines a bipartite graph G with L.G/ D A, R.G/ D
b G. Exhibit a subset S of A such that both
B and an edge between F and G iff F R
S and A S are nonempty, and the set N.S / of neighbors of S is the same size as
S , that is, jN.S /j D jS j.
(d) Let G be an arbitrary, finite, bipartite graph. For any subset S L.G/, let
S WWD L.G/ S , and likewise for any M R.G/, let M WWD R.G/ M . Suppose
S is a subset of L.G/ such that jN.S /j D jS j, and both S and S are nonempty.
Circle the formula that correctly completes the following statement:
There is a matching from L.G/ to R.G/ if and only if there is both a matching
from S to its neighbors, N.S /, and also a matching from S to
N.S /
N.S /
N
1
.N.S //
N
1
.N.S //
N.S / N.S /
N.S / N.S /
Hint: The proof of Hall’s Bottleneck Theorem.
Problem 12.17. (a) Show that there is no matching for the bipartite graph G in
Figure 12.25 that covers L.G/.
(b) The bipartite graph H in Figure 12.26 has an easily verified property that
implies it has a matching that covers L.H /. What is the property?
“mcs” — 2017/6/5 — 19:42 — page 504 — #512
504
Chapter 12
Simple Graphs
a
v
b
w
c
x
d
y
e
z
L(G)
Figure 12.25
R(G)
Bipartite graph G.
Homework Problems
Problem 12.18.
A Latin square is n n array whose entries are the number 1; : : : ; n. These entries satisfy two constraints: every row contains all n integers in some order, and
also every column contains all n integers in some order. Latin squares come up
frequently in the design of scientific experiments for reasons illustrated by a little
story in a footnote10
10 At Guinness brewery in the eary 1900’s, W. S. Gosset (a chemist) and E. S. Beavan (a “maltster”)
were trying to improve the barley used to make the brew. The brewery used different varieties of
barley according to price and availability, and their agricultural consultants suggested a different
fertilizer mix and best planting month for each variety.
Somewhat sceptical about paying high prices for customized fertilizer, Gosset and Beavan planned
a season long test of the influence of fertilizer and planting month on barley yields. For as many
months as there were varieties of barley, they would plant one sample of each variety using a different
one of the fertilizers. So every month, they would have all the barley varieties planted and all the
fertilizers used, which would give them a way to judge the overall quality of that planting month.
But they also wanted to judge the fertilizers, so they wanted each fertilizer to be used on each variety
during the course of the season. Now they had a little mathematical problem, which we can abstract
as follows.
Suppose there are n barley varieties and an equal number of recommended fertilizers. Form an
n n array with a column for each fertilizer and a row for each planting month. We want to fill in
the entries of this array with the integers 1,. . . ,n numbering the barley varieties, so that every row
contains all n integers in some order (so every month each variety is planted and each fertilizer is
used), and also every column contains all n integers (so each fertilizer is used on all the varieties over
the course of the growing season).
“mcs” — 2017/6/5 — 19:42 — page 505 — #513
12.12. References
505
v
a
w
b
x
c
y
d
z
L(H)
Figure 12.26
R(H)
Bipartite Graph H .
For example, here is a 4 4 Latin square:
1
3
2
4
2
4
1
3
3
2
4
1
4
1
3
2
(a) Here are three rows of what could be part of a 5 5 Latin square:
2 4 5 3 1
4 1 3 2 5
3 2 1 5 4
Fill in the last two rows to extend this “Latin rectangle” to a complete Latin square.
(b) Show that filling in the next row of an n n Latin rectangle is equivalent to
finding a matching in some 2n-vertex bipartite graph.
(c) Prove that a matching must exist in this bipartite graph and, consequently, a
Latin rectangle can always be extended to a Latin square.
“mcs” — 2017/6/5 — 19:42 — page 506 — #514
506
Chapter 12
Simple Graphs
Problem 12.19.
Take a regular deck of 52 cards. Each card has a suit and a value. The suit is one of
four possibilities: heart, diamond, club, spade. The value is one of 13 possibilities,
A; 2; 3; : : : ; 10; J; Q; K. There is exactly one card for each of the 4 13 possible
combinations of suit and value.
Ask your friend to lay the cards out into a grid with 4 rows and 13 columns.
They can fill the cards in any way they’d like. In this problem you will show that
you can always pick out 13 cards, one from each column of the grid, so that you
wind up with cards of all 13 possible values.
(a) Explain how to model this trick as a bipartite matching problem between the
13 column vertices and the 13 value vertices. Is the graph necessarily degreeconstrained?
(b) Show that any n columns must contain at least n different values and prove
that a matching must exist.
Problem 12.20.
Scholars through the ages have identified twenty fundamental human virtues: honesty, generosity, loyalty, prudence, completing the weekly course reading-response,
etc. At the beginning of the term, every student in Math for Computer Science possessed exactly eight of these virtues. Furthermore, every student was unique; that
is, no two students possessed exactly the same set of virtues. The Math for Computer Science course staff must select one additional virtue to impart to each student
by the end of the term. Prove that there is a way to select an additional virtue for
each student so that every student is unique at the end of the term as well.
Suggestion: Use Hall’s theorem. Try various interpretations for the vertices on
the left and right sides of your bipartite graph.
Problem 12.21.
Suppose 2m teams play in a round-robin tournament. Over a period of 2m 1
days, every team plays every other team exactly once. There are no ties. Show that
for each day we can select a winning team, without selecting the same team twice.
“mcs” — 2017/6/5 — 19:42 — page 507 — #515
12.12. References
507
Problems for Section 12.6
Class Problems
Problem 12.22.
Let G be the graph below.11 Carefully explain why .G/ D 4.
Problem 12.23.
A portion of a computer program consists of a sequence of calculations where the
results are stored in variables, like this:
Step 1:
2:
3:
4:
5:
6:
Inputs:
c
d
e
f
g
h
Outputs:
D
D
D
D
D
D
a; b
aCb
ac
cC3
c e
aCf
f C1
d; g; h
A computer can perform such calculations most quickly if the value of each variable
is stored in a register, a chunk of very fast memory inside the microprocessor.
Programming language compilers face the problem of assigning each variable in a
program to a register. Computers usually have few registers, however, so they must
be used wisely and reused often. This is called the register allocation problem.
In the example above, variables a and b must be assigned different registers,
because they hold distinct input values. Furthermore, c and d must be assigned
different registers; if they used the same one, then the value of c would be overwritten in the second step and we’d get the wrong answer in the third step. On the
11 From
[32], Exercise 13.3.1
“mcs” — 2017/6/5 — 19:42 — page 508 — #516
508
Chapter 12
Simple Graphs
other hand, variables b and d may use the same register; after the first step, we no
longer need b and can overwrite the register that holds its value. Also, f and h may
use the same register; once f C 1 is evaluated in the last step, the register holding
the value of f can be overwritten.
(a) Recast the register allocation problem as a question about graph coloring.
What do the vertices correspond to? Under what conditions should there be an edge
between two vertices? Construct the graph corresponding to the example above.
(b) Color your graph using as few colors as you can. Call the computer’s registers
R1, R2 etc. Describe the assignment of variables to registers implied by your
coloring. How many registers do you need?
(c) Suppose that a variable is assigned a value more than once, as in the code
snippet below:
:::
t Dr Cs
uDt 3
t Dm
k
v Dt Cu
:::
How might you cope with this complication?
Problem 12.24.
Suppose an n-vertex bipartite graph has exactly k connected components, each of
which has two or more vertices. How many ways are there color it using a given
set of two colors?
Homework Problems
Problem 12.25.
6.042 is often taught using recitations. Suppose it happened that 8 recitations were
needed, with two or three staff members running each recitation. The assignment
of staff to recitation sections, using their secret codenames, is as follows:
R1: Maverick, Goose, Iceman
R2: Maverick, Stinger, Viper
“mcs” — 2017/6/5 — 19:42 — page 509 — #517
12.12. References
509
R3: Goose, Merlin
R4: Slider, Stinger, Cougar
R5: Slider, Jester, Viper
R6: Jester, Merlin
R7: Jester, Stinger
R8: Goose, Merlin, Viper
Two recitations can not be held in the same 90-minute time slot if some staff
member is assigned to both recitations. The problem is to determine the minimum
number of time slots required to complete all the recitations.
(a) Recast this problem as a question about coloring the vertices of a particular
graph. Draw the graph and explain what the vertices, edges, and colors represent.
(b) Show a coloring of this graph using the fewest possible colors; explain why
no fewer colors will work. What schedule of recitations does this imply?
Problem 12.26.
This problem generalizes the result proved Theorem 12.6.3 that any graph with
maximum degree at most w is .w C 1/-colorable.
A simple graph G is said to have width w iff its vertices can be arranged in a
sequence such that each vertex is adjacent to at most w vertices that precede it in
the sequence. If the degree of every vertex is at most w, then the graph obviously
has width at most w—just list the vertices in any order.
(a) Prove that every graph with width at most w is .w C 1/-colorable.
(b) Describe a 2-colorable graph with minimum width n.
(c) Prove that the average degree of a graph of width w is at most 2w.
(d) Describe an example of a graph with 100 vertices, width 3, but average degree
more than 5.
Problem 12.27.
A sequence of vertices of a graph has width w iff each vertex is adjacent to at most
w vertices that precede it in the sequence. A simple graph G has width w if there
is a width-w sequence of all its vertices.
“mcs” — 2017/6/5 — 19:42 — page 510 — #518
510
Chapter 12
Simple Graphs
(a) Explain why the width of a graph must be at least the minimum degree of its
vertices.
(b) Prove that if a finite graph has width w, then there is a width-w sequence of
all its vertices that ends with a minimum degree vertex.
(c) Describe a simple algorithm to find the minimum width of a graph.
Problem 12.28.
Let G be a simple graph whose vertex degrees are all k. Prove by induction on
number of vertices that if every connected component of G has a vertex of degree
strictly less than k, then G is k-colorable.
Problem 12.29.
A basic example of a simple graph with chromatic number n is the complete graph
on n vertices, that is .Kn / D n. This implies that any graph with Kn as a subgraph
must have chromatic number at least n. It’s a common misconception to think that,
conversely, graphs with high chromatic number must contain a large complete subgraph. In this problem we exhibit a simple example countering this misconception,
namely a graph with chromatic number four that contains no triangle—length three
cycle—and hence no subgraph isomorphic to Kn for n 3. Namely, let G be the
11-vertex graph of Figure 12.27. The reader can verify that G is triangle-free.
Figure 12.27
Graph G with no triangles and .G/ D 4.
(a) Show that G is 4-colorable.
(b) Prove that G can’t be colored with 3 colors.
“mcs” — 2017/6/5 — 19:42 — page 511 — #519
12.12. References
511
Problem 12.30.
This problem will show that 3-coloring a graph is just as difficult as finding a satisfying truth assignment for a propositional formula. The graphs considered will
all be taken to have three designated color-vertices connected in a triangle to force
them to have different colors in any coloring of the graph. The colors assigned to
the color-vertices will be called T; F and N .
Suppose f is an n-argument truth function. That is,
f W fT; F gn ! fT; F g:
A graph G is called a 3-color-f-gate iff G has n designated input vertices and a
designated output vertex, such that
G can be 3-colored only if its input vertices are colored with T ’s and F ’s.
For every sequence b1 ; b2 ; : : : ; bn 2 fT; F g, there is a 3-coloring of G in
which the input vertices v1 ; v2 ; : : : ; vn 2 V .G/ have the colors b1 ; b2 ; : : : ; bn 2
fT; F g.
In any 3-coloring of G where the input vertices v1 ; v2 ; : : : ; vn 2 V .G/ have
colors b1 ; b2 ; : : : ; bn 2 fT; F g, the output vertex has color f .b1 ; b2 ; : : : ; bn /.
For example, a 3-color-NOT-gate consists simply of two adjacent vertices. One
vertex is designated to be the input vertex P and the other is designated to be the
output vertex. Both vertices have to be constrained so they can only be colored with
T ’s or F ’s in any proper 3-coloring. This constraint can be imposed by making
them adjacent to the color-vertex N , as shown in Figure 12.28.
(a) Verify that the graph in Figure 12.29 is a 3-color-OR-gate. (The dotted lines
indicate edges to color-vertex N ; these edges constrain the P , Q and P OR Q
vertices to be colored T or F in any proper 3-coloring.)
(b) Let E be an n-variable propositional formula, and suppose E defines a truth
function f W fT; F gn ! fT; F g. Explain a simple way to construct a graph that is
a 3-color-f -gate.
(c) Explain why an efficient procedure for determining if a graph was 3-colorable
would lead to an efficient procedure to solve the satisfiability problem, SAT.
Problem 12.31.
The 3-coloring problem for planar graphs turns out to be no easier than the 3coloring problem for arbitrary graphs. This claim follows very simply from the
“mcs” — 2017/6/5 — 19:42 — page 512 — #520
512
Chapter 12
Simple Graphs
NOT(P)
P
N T F [h]
Figure 12.28 A 3-color NOT-gate
P OR Q
N
T
F
P
[h]
Figure 12.29 A 3-color OR-gate
Q
“mcs” — 2017/6/5 — 19:42 — page 513 — #521
12.12. References
513
u v x w [h]
Figure 12.30 A 3-color cross-over gadget.
existence of a “3-color cross-over gadget.” Such a gadget is a planar graph whose
outer face is a cycle with four designated vertices u; v; w; x occurring in clockwise
order such that
1. Any assignment of colors to vertices u and v can be completed into a 3coloring of the gadget.
2. In every 3-coloring of the gadget, the colors of u and w are the same, and the
colors of v and x are the also same.
Figure 12.30 shows such a 3-color cross-over gadget.12
So to find a 3-coloring for any simple graph, simply draw it in the plane with
edges crossing as needed, and then replace each occurrence of an edge crossing by
a copy of the gadget as shown in Figure 12.31. This yields a planar graph which
has a 3-coloring iff the original graph had one.
(a) Prove that the graph in Figure 12.30 satisfies condition (1) by exhibiting the
claimed 3-colorings.
Hint: Only two colorings are needed, one where u and v are the same color and
another where they are not the same color.
12 This
gadget and reduction of 3-colorability to planar 3-colorability are due to Larry Stockmeyer [46].
“mcs” — 2017/6/5 — 19:42 — page 514 — #522
514
Chapter 12
Simple Graphs
[h]
Figure 12.31
Replacing an edge-crossing with a planar gadget.
(b) Prove that the graph in Figure 12.30 satisfies condition (2).
Hint: The colorings for part (a) are almost completely forced by the coloring of u
and v.
Exam Problems
Problem 12.32.
False Claim. Let G be a graph whose vertex degrees are all k. If G has a vertex
of degree strictly less than k, then G is k-colorable.
(a) Give a counterexample to the False Claim when k D 2.
(b) Underline the exact sentence or part of a sentence that is the first unjustified
step in the following bogus proof of the False Claim.
Bogus proof. Proof by induction on the number n of vertices:
The induction hypothesis P .n/ is:
Let G be an n-vertex graph whose vertex degrees are all k. If G
also has a vertex of degree strictly less than k, then G is k-colorable.
Base case: (n D 1) G has one vertex, the degree of which is 0. Since G is
1-colorable, P .1/ holds.
Inductive step: We may assume P .n/. To prove P .n C 1/, let GnC1 be a
graph with nC1 vertices whose vertex degrees are all k or less. Also, suppose
GnC1 has a vertex v of degree strictly less than k. Now we only need to prove
that GnC1 is k-colorable.
“mcs” — 2017/6/5 — 19:42 — page 515 — #523
12.12. References
515
To do this, first remove the vertex v to produce a graph Gn with n vertices.
Let u be a vertex that is adjacent to v in GnC1 . Removing v reduces the
degree of u by 1. So in Gn , vertex u has degree strictly less than k. Since no
edges were added, the vertex degrees of Gn remain k. So Gn satisfies the
conditions of the induction hypothesis P .n/, and so we conclude that Gn is
k-colorable.
Now a k-coloring of Gn gives a coloring of all the vertices of GnC1 , except for
v. Since v has degree less than k, there will be fewer than k colors assigned
to the nodes adjacent to v. So among the k possible colors, there will be a
color not used to color these adjacent nodes, and this color can be assigned to
v to form a k-coloring of GnC1 .
(c) With a slightly strengthened condition, the preceding proof of the False Claim
could be revised into a sound proof of the following Claim:
Claim. Let G be a graph whose vertex degrees are all k.
If hstatement inserted from belowi has a vertex of degree strictly less than k, then
G is k-colorable.
Indicate each of the statements below that could be inserted to make the proof
correct.
G is connected and
G has no vertex of degree zero and
G does not contain a complete graph on k vertices and
every connected component of G
some connected component of G
Problem 12.33.
In the graph shown in Figure 12.32, the vertices connected in the triangle on the left
are called color-vertices; since they form a triangle, they are forced to have different
colors in any coloring of the graph. The colors assigned to the color-vertices will
be called T; F and N. The dotted lines indicate edges to the color-vertex N.
(a) Explain why for any assignment of different truth-colors to P and Q, there is
a unique 3-coloring of the graph.
(b) Prove that in any 3-coloring of the whole graph, the vertex labeled P XOR Q
is colored with the XOR of the colors of vertices P and Q.
“mcs” — 2017/6/5 — 19:42 — page 516 — #524
516
Chapter 12
Simple Graphs
XOR
P ⊕ Q
N
b
T
e
P
c
F
F
d
a
Q
[h]
Figure 12.32 A 3-color XOR-gate
Problems for Section 12.7
Exam Problems
Problem 12.34.
Since you can go back and forth on an edge in a simple graph, every vertex is on
an even length closed walk. So even length closed walks don’t tell you much about
even length cycles. The situation with odd-length closed walks is more interesting.
(a) Give an example of a simple graph in which every vertex is on a unique oddlength cycle and a unique even-length cycle.
Hint: Four vertices.
(b) Give an example of a simple graph in which every vertex is on a unique oddlength cycle and no vertex is on an even-length cycle.
(c) Prove that in a digraph, a smallest size odd-length closed walk must be a cycle.
Note that there will always be lots of even-length closed walks that are shorter than
the smallest odd-length one.
Hint: Let e be an odd-length closed walk of minimum size, and suppose it begins
and ends at vertex a. If it is not a cycle, then it must include a repeated vertex
“mcs” — 2017/6/5 — 19:42 — page 517 — #525
12.12. References
517
b ¤ a. That is, e starts with a walk f from a to b, followed by a walk g from b to
b, followed by a walk h from b to a.13
Homework Problems
Problem 12.35. (a) Give an example of a simple graph that has two vertices u ¤ v
and two distinct paths between u and v, but neither u nor v is on a cycle.
(b) Prove that if there are different paths between two vertices in a simple graph,
then the graph has a cycle.
Problems for Section 12.10
Class Problems
Problem 12.36.
A simple graph G is 2-removable iff it contains two vertices v ¤ w such that G v
is connected, and G w is also connected. Prove that every connected graph with
at least two vertices is 2-removable.
Hint: Consider a maximum length path.
Problem 12.37.
The n-dimensional hypercube Hn is a graph whose vertices are the binary strings
of length n. Two vertices are adjacent if and only if they differ in exactly 1 bit. For
example, in H3 , vertices 111 and 011 are adjacent because they differ only in the
first bit, while vertices 101 and 011 are not adjacent because they differ at both
the first and second bits.
(a) Prove that it is impossible to find two spanning trees of H3 that do not share
some edge.
(b) Verify that for any two vertices x ¤ y of H3 , there are 3 paths from x to y in
H3 , such that, besides x and y, no two of those paths have a vertex in common.
(c) Conclude that the connectivity of H3 is 3.
13 In
the notation of the text
e D a fb
b gb
b h a:
“mcs” — 2017/6/5 — 19:42 — page 518 — #526
518
Chapter 12
Simple Graphs
(d) Try extending your reasoning to H4 . (In fact, the connectivity of Hn is n for
all n 1. A proof appears in the problem solution.)
Problem 12.38.
A set M of vertices of a graph is a maximal connected set if every pair of vertices
in the set are connected, and any set of vertices properly containing M will contain
two vertices that are not connected.
(a) What are the maximal connected subsets of the following (unconnected) graph?
(b) Explain the connection between maximal connected sets and connected components. Prove it.
Problem 12.39. (a) Prove that Kn is .n 1/-edge connected for n > 1.
Let Mn be a graph defined as follows: begin by taking n graphs with nonoverlapping sets of vertices, where each of the n graphs is .n 1/-edge connected
(they could be disjoint copies of Kn , for example). These will be subgraphs of Mn .
Then pick n vertices, one from each subgraph, and add enough edges between pairs
of picked vertices that the subgraph of the n picked vertices is also .n 1/-edge
connected.
(b) Draw a picture of M3 .: : : M4 /.
(c) Explain why Mn is .n
Problem 12.40.
1/-edge connected.
“mcs” — 2017/6/5 — 19:42 — page 519 — #527
12.12. References
519
False Claim. If every vertex in a graph has positive degree, then the graph is
connected.
(a) Prove that this Claim is indeed false by providing a counterexample.
(b) Since the Claim is false, there must be a logical mistake in the following bogus
proof. Pinpoint the first logical mistake (unjustified step) in the proof.
Bogus proof. We prove the Claim above by induction. Let P .n/ be the proposition
that if every vertex in an n-vertex graph has positive degree, then the graph is
connected.
Base cases: (n 2). In a graph with 1 vertex, that vertex cannot have positive
degree, so P .1/ holds vacuously.
P .2/ holds because there is only one graph with two vertices of positive degree,
namely, the graph with an edge between the vertices, and this graph is connected.
Inductive step: We must show that P .n/ implies P .n C 1/ for all n 2. Consider
an n-vertex graph in which every vertex has positive degree. By the assumption
P .n/, this graph is connected; that is, there is a path between every pair of vertices.
Now we add one more vertex x to obtain an .n C 1/-vertex graph:
z
n-node
connected
graph
x
y
All that remains is to check that there is a path from x to every other vertex z. Since
x has positive degree, there is an edge from x to some other vertex y. Thus, we can
obtain a path from x to z by going from x to y and then following the path from y
to z. This proves P .n C 1/.
By the principle of induction, P .n/ is true for all n 0, which proves the Claim.
“mcs” — 2017/6/5 — 19:42 — page 520 — #528
520
Chapter 12
Simple Graphs
Homework Problems
Problem 12.41.
An edge is said to leave a set of vertices if one end of the edge is in the set and the
other end is not.
(a) An n-node graph is said to be mangled if there is an edge leaving every set of
bn=2c or fewer vertices. Prove the following:
Claim. Every mangled graph is connected.
An n-node graph is said to be tangled if there is an edge leaving every set of
dn=3e or fewer vertices.
(b) Draw a tangled graph that is not connected.
(c) Find the error in the bogus proof of the following
False Claim. Every tangled graph is connected.
Bogus proof. The proof is by strong induction on the number of vertices in the
graph. Let P .n/ be the proposition that if an n-node graph is tangled, then it is
connected. In the base case, P .1/ is true because the graph consisting of a single
node is trivially connected.
For the inductive case, assume n 1 and P .1/; : : : ; P .n/ hold. We must prove
P .n C 1/, namely, that if an .n C 1/-node graph is tangled, then it is connected.
So let G be a tangled, .n C 1/-node graph. Choose dn=3e of the vertices and let G1
be the tangled subgraph of G with these vertices and G2 be the tangled subgraph
with the rest of the vertices. Note that since n 1, the graph G has a least two
vertices, and so both G1 and G2 contain at least one vertex. Since G1 and G2 are
tangled, we may assume by strong induction that both are connected. Also, since
G is tangled, there is an edge leaving the vertices of G1 which necessarily connects
to a vertex of G2 . This means there is a path between any two vertices of G: a path
within one subgraph if both vertices are in the same subgraph, and a path traversing
the connecting edge if the vertices are in separate subgraphs. Therefore, the entire
graph G is connected. This completes the proof of the inductive case, and the Claim
follows by strong induction.
Problem 12.42.
In the cycle C2n of length 2n, we’ll call two vertices opposite if they are on opposite
sides of the cycle, that is that are distance n apart in C2n . Let G be the graph formed
“mcs” — 2017/6/5 — 19:42 — page 521 — #529
12.12. References
521
from C2n by adding an edge, which we’ll call a crossing edge, between each pair
of opposite vertices. So G has n crossing edges.
(a) Give a simple description of the shortest path between any two vertices of G.
Hint: Argue that a shortest path between two vertices in G uses at most one crossing
edge.
(b) What is the diameter of G, that is, the largest distance between two vertices?
(c) Prove that the graph is not 4-connected.
(d) Prove that the graph is 3-connected.
Problem 12.43.
An Euler tour of a graph is a closed walk that includes every edge exactly once.14
In this problem we work out a proof of:
Theorem. A connected graph has an Euler tour if and only if every vertex has even
degree.
(a) Show that if a graph has an Euler tour, then the degree of each of its vertices
is even.
In the remaining parts, we’ll work out the converse: if the degree of every vertex
of a connected finite graph is even, then it has an Euler tour. To do this, let’s define
an Euler walk to be a walk that includes each edge at most once.
(b) Suppose that an Euler walk in a connected graph does not include every edge.
Explain why there must be an unincluded edge that is incident to a vertex on the
walk.
In the remaining parts, let w be the longest Euler walk in some finite, connected
graph.
(c) Show that if w is a closed walk, then it must be an Euler tour.
Hint: part (b)
(d) Explain why all the edges incident to the end of w must already be in w.
(e) Show that if the end of w was not equal to the start of w, then the degree of
the end would be odd.
Hint: part (d)
14 In
some other texts, this is called an Euler circuit.
“mcs” — 2017/6/5 — 19:42 — page 522 — #530
522
Chapter 12
Simple Graphs
(f) Conclude that if every vertex of a finite, connected graph has even degree, then
it has an Euler tour.
Exam Problems
Problem 12.44.
We apply the following operation to a simple graph G: pick two vertices u ¤ v
such that either
1. there is an edge of G between u and v, and there is also a path from u to v
which does not include this edge; in this case, delete the edge hu—vi.
2. there is no path from u to v; in this case, add the edge hu—vi.
Keep repeating these operations until it is no longer possible to find two vertices
u ¤ v to which an operation applies.
Assume the vertices of G are the integers 1; 2; : : : ; n for some n 2. This
procedure can be modelled as a state machine whose states are all possible simple
graphs with vertices 1; 2; : : : ; n. G is the start state, and the final states are the
graphs on which no operation is possible.
(a) Let G be the graph with vertices f1; 2; 3; 4g and edges
ff1; 2g; f3; 4gg
How many possible final states are reachable from start state G?
1in
(b) On the line next to each of the derived state variables below, indicate the
strongest property from the list below that the variable is guaranteed to satisfy,
no matter what the starting graph G is. The properties are:
constant
nonincreasing
increasing
nondecreasing
decreasing
none of these
For any state, let e be the number of edges in it, and let c be the number of connected components it has. Since e may increase or decrease in a transition, it does
not have any of the first four properties. The derived variables are:
0) e
none of these
i) c
ii) c C e
iii) 2c C e
iv) c C
e
eC1
“mcs” — 2017/6/5 — 19:42 — page 523 — #531
12.12. References
523
(c) Explain why, starting from any state G, the procedure terminates. If your explanation depends on answers you gave to part (b), you must justify those answers.
(d) Prove that any final state must be an unordered tree on the set of vertices, that
is, a spanning tree.
Problem 12.45.
If a simple graph has e edges, v vertices, and k connected components, then it has
at least e v C k cycles.
Prove this by induction on the number of edges e.
Problems for Section 12.11
Practice Problems
Problem 12.46. (a) Prove that the average degree of a tree is less than 2.
(b) Suppose every vertex in a graph has degree at least k. Explain why the graph
has a path of length k.
Hint: Consider a longest path.
a
c
d
b
e
g
f
h
Figure 12.33
The graph G.
“mcs” — 2017/6/5 — 19:42 — page 524 — #532
524
Chapter 12
Simple Graphs
Problem 12.47. (a) How many spanning trees are there for the graph G in Figure 12.33?
(b) For G e, the graph G with vertex e deleted, describe two spanning trees that
have no edges in common.
(c) For G e with edge ha—d i deleted, explain why there cannot be two edgedisjoint spanning trees.
Hint: : Count vertices and edges.
Problem 12.48.
Prove that if G is a forest and
jV .G/j D jE.G/j C 1;
(12.6)
then G is a tree.
Problem 12.49.
Let H3 be the graph shown in Figure 12.34. Explain why it is impossible to find
two spanning trees of H3 that have no edges in common.
010
000
001
100
110
101
111
011
Figure 12.34 H3 .
Problem 12.50.
The diameter of a connected graph is the largest distance between any two vertices.
(a) What is the largest possible diameter in any connected graph with n vertices?
Describe a graph with this maximum diameter.
(b) What is the smallest possible diameter of an n-vertex tree for n > 2? Describe
an n-vertex tree with this minimum diameter.
“mcs” — 2017/6/5 — 19:42 — page 525 — #533
12.12. References
525
Problem 12.51.
(a) Indicate all the properties below that are preserved under graph isomorphism.
There is a cycle that includes all the vertices.
Two edges are of equal length.
The graph remains connected if any two edges are removed.
There exists an edge that is an edge of every spanning tree.
The negation of a property that is preserved under isomorphism.
(b) For the following statements about finite trees, indicate whether they are true
or false, and provide counterexamples for those that are false.
Any connected subgraph is a tree.
true
false
Adding an edge between two nonadjacent vertices creates a cycle.
false
true
The number of vertices is one less than twice the number of leaves.
false
true
The number of vertices is one less than the number of edges. true
false
For every finite graph (not necessarily a tree), there is one (a finite tree) that
spans it.
true
false
Problem 12.52.
Circle true or false for the following statements about finite simple
graphs G.
(a) G has a spanning tree.
true
false
(b) jV .G/j D O.jE.G/j/ for connected G.
true
false
(c) jE.G/j D O.jV .G/j/.
true
false
(d) .G/ maxfdeg.v/ j v 2 V .G/g.15
true
false
(e) jV .G/j D O..G//.
true
false
15 .G/
is the chromatic number of G.
“mcs” — 2017/6/5 — 19:42 — page 526 — #534
526
Chapter 12
Simple Graphs
Exam Problems
Problem 12.53. (a) Let T be a tree and e a new edge between two vertices of T .
Explain why T C e must contain a cycle.
(b) Conclude that T C e must have at least three spanning trees.
Problem 12.54.
How many 5-vertex non-isomomorphic trees are there? Explain.
Problem 12.55.
A simple graph G is said to have width 1 iff there is a way to list all its vertices so
that each vertex is adjacent to at most one vertex that appears earlier in the list. All
the graphs mentioned below are assumed to be finite.
(a) Prove that every graph with width one is a forest.
Hint: By induction, removing the last vertex.
(b) Prove that every finite tree has width one. Conclude that a graph is a forest iff
it has width one.
Problem 12.56.
Prove by induction that, using a fixed set of n > 1 colors, there are exactly
n .n
1/m
1
different colorings of any tree with m vertices.
Problem 12.57.
Let G be a connected weighted simple graph and let v be a vertex of G. Suppose
e WWD hv—wi is an edge of G that is strictly smaller than the weight of every other
edge incident to v. Let T be a minimum weight spanning tree of G. Prove that e is
an edge of T . Hint: By contradiction.
Problem 12.58.
Note that the two parts of this problem do not depend on each other.
“mcs” — 2017/6/5 — 19:42 — page 527 — #535
12.12. References
527
(a) Let G be a connected simple graph. Prove that every spanning tree of G must
include every cut edge of G.
(b) Suppose a connected, weighted graph G has a unique maximum-weight edge
e. Show that if e is in a minimum weight spanning tree of G, then e is a cut edge.
Class Problems
Problem 12.59.
Procedure Mark starts with a connected, simple graph with all edges unmarked and
then marks some edges. At any point in the procedure a path that includes only
marked edges is called a fully marked path, and an edge that has no fully marked
path between its endpoints is called eligible.
Procedure Mark simply keeps marking eligible edges, and terminates when there
are none.
Prove that Mark terminates, and that when it does, the set of marked edges forms
a spanning tree of the original graph.
Problem 12.60.
A procedure for connecting up a (possibly disconnected) simple graph and creating
a spanning tree can be modelled as a state machine whose states are finite simple
graphs. A state is final when no further transitions are possible. The transitions are
determined by the following rules:
Procedure create-spanning-tree
1. If there is an edge hu—vi on a cycle, then delete hu—vi.
2. If vertices u and v are not connected, then add the edge hu—vi.
(a) Draw all the possible final states reachable starting with the graph with vertices
f1; 2; 3; 4g and edges
fh1—2i ; h3—4ig:
(b) Prove that if the machine reaches a final state, then the final state will be a tree
on the vertices of the graph on which it started.
(c) For any graph G 0 , let e be the number of edges in G 0 , c be the number of connected components it has, and s be the number of cycles. For each of the quantities
“mcs” — 2017/6/5 — 19:42 — page 528 — #536
528
Chapter 12
Simple Graphs
below, indicate the strongest of the properties that it is guaranteed to satisfy, no
matter what the starting graph is.
The choices for properties are: constant, strictly increasing, strictly decreasing,
weakly increasing, weakly decreasing, none of these.
(i) e
(ii) c
(iii) s
(iv) e
s
(v) c C e
(vi) 3c C 2e
(vii) c C s
(d) Prove that one of the quantities from part (c) strictly decreases at each transition. Conclude that for every starting state, the machine will reach a final state.
Problem 12.61.
Let G be a weighted graph and suppose there is a unique edge e 2 E.G/ with
smallest weight, that is, w.e/ < w.f / for all edges f 2 E.G/ feg. Prove that
any minimum weight spanning tree (MST) of G must include e.
Problem 12.62.
Let G be the 4 4 grid with vertical and horizontal edges between neighboring
vertices and edge weights as shown in Figure 12.35.
In this problem you will practice some of the ways to build minimum-weight
spanning trees. For each part, list the edge weights in the order in which the edges
with those weights were chosen by the given rules.
(a) Construct a minimum weight spanning tree (MST) for G by initially selecting
the minimum weight edge, and then successively selecting the minimum weight
edge that does not create a cycle with the previously selected edges. Stop when the
selected edges form a spanning tree of G. (This is Kruskal’s MST algorithm.)
For any step in Kruskal’s procedure, describe a black-white coloring of the graph
components so that the edge Kruskal chooses is the minimum weight “gray edge”
according to Lemma 12.11.8.
“mcs” — 2017/6/5 — 19:42 — page 529 — #537
12.12. References
529
0.04
0
1
1.04
0.01
1.01
u
1.02
0.05
0.09
1.09
0.06
1.06
0.03
v
Figure 12.35
1.12
1.08
1.05
0.02
0.08
w
1.13
0.10
1.10
0.07
1.14
0.11
The 4x4 array graph G
(b) Grow an MST for G by starting with the tree consisting of the single vertex u
and then successively adding the minimum weight edge with exactly one endpoint
in the tree. Stop when the tree spans G. (This is Prim’s MST algorithm.)
For any step in Prim’s procedure, describe a black-white coloring of the graph components so that the edge Prim chooses is the minimum weight “gray edge”according
to Lemma 12.11.8.
(c) The 6.042 “parallel” MST algorithm can grow an MST for G by starting with
the upper left corner vertex along with the vertices labelled v and w. Regard each of
the three vertices as one-vertex trees. Successively add, for each tree in parallel, the
minimum-weight edge among the edges with exactly one endpoint in the tree. Stop
working on a tree when there is an edge connecting it to another tree. Continue
until there are no more eligible trees—that is, each tree is connected by an edge
to another tree—then go back to applying the general gray-edge method until the
parallel trees merge to form a spanning tree of G.
“mcs” — 2017/6/5 — 19:42 — page 530 — #538
530
Chapter 12
Simple Graphs
(d) Verify that you got the same MST each time as promised by Corollary 12.11.11.
Problem 12.63.
In this problem you will prove:
Theorem. A graph G is 2-colorable iff it contains no odd length closed walk.
As usual with “iff” assertions, the proof splits into two proofs: part (a) asks you
to prove that the left side of the “iff” implies the right side. The other problem parts
prove that the right side implies the left.
(a) Assume the left side and prove the right side. Three to five sentences should
suffice.
(b) Now assume the right side. As a first step toward proving the left side, explain
why we can focus on a single connected component H within G.
(c) As a second step, explain how to 2-color any tree.
(d) Choose any 2-coloring of a spanning tree T of H . Prove that H is 2-colorable
by showing that any edge not in T must also connect different-colored vertices.
Problem 12.64.
MIT Information Services & Technology (IS&T) wants to assemble a cluster of
n computers using wires and hubs. Each computer must have exactly one wire
attached to it, while each hub can have up to five wires attached. There must be a
path of wires between every pair of computers in the cluster.
(a) Prove by induction that d.n
the cluster of n computers.
2/=3e hubs are sufficient for IS&T to assemble
Suppose IS&T uses m hubs.
(b) Write a simple formula in terms of m and n that bounds the maximum number
of wires that can be attached to the hubs and computers without leaving dangling
ends.
(c) Write a simple formula in terms of m and n for a lower bound on the number
of wires needed to form the cluster, even using hubs that may have more than five
wires attached.
“mcs” — 2017/6/5 — 19:42 — page 531 — #539
12.12. References
531
(d) IS&T wants to minimize the number of hubs used to assemble the cluster. Use
the previous results to prove that at d.n 2/=3e hubs are necessary and sufficient
to hook up the cluster.
Homework Problems
Problem 12.65.
Suppose D D hd1 ; d2 ; : : : ; dn i is a list of the degrees of the vertices of some nvertex tree T for n 2. That is, we assume the vertices of T are numbered, and
di > 0 is the degree of the i th vertex of T .
(a) Explain why
n
X
di D 2.n
1/:
(12.7)
i D1
Say a finite sequence of n 2 positive integers is good when the sum of its
elements is 2.n 1/.
(b) Prove that every good sequence is a list of the degrees of all the vertices in
some tree.
Hint: Induction.
(c) Show that if D is a good sequence, then the elements of D can each be assigned one of two colors, say “red” and “black,” such that the sum of the red elements equals the sum of the black elements.
Problem 12.66.
Prove Corollary 12.11.11: If all edges in a finite weighted graph have distinct
weights, then the graph has a unique MST.
Hint: Suppose M and N were different MST’s of the same graph. Let e be the
smallest edge in one and not the other, say e 2 M N , and observe that N C e
must have a cycle.
“mcs” — 2017/6/5 — 19:42 — page 532 — #540
“mcs” — 2017/6/5 — 19:42 — page 533 — #541
13
Planar Graphs
13.1
Drawing Graphs in the Plane
Suppose there are three dog houses and three human houses, as shown in Figure 13.1. Can you find a route from each dog house to each human house such that
no route crosses any other route?
A similar question comes up about a little-known animal called a quadrapus that
looks like an octopus with four stretchy arms instead of eight. If five quadrapi are
resting on the sea floor, as shown in Figure 13.2, can each quadrapus simultaneously shake hands with every other in such a way that no arms cross?
Both these puzzles can be understood as asking about drawing graphs in the
plane. Replacing dogs and houses by nodes, the dog house puzzle can be rephrased
as asking whether there is a planar drawing of the graph with six nodes and edges
between each of the first three nodes and each of the second three nodes. This
graph is called the complete bipartite graph K3;3 and is shown in Figure 13.3.(a).
The quadrapi puzzle asks whether there is a planar drawing of the complete graph
K5 shown in Figure 13.3.(b).
In each case, the answer is, “No—but almost!” In fact, if you remove an edge
from either of these graphs, then the resulting graph can be redrawn in the plane so
that no edges cross, as shown in Figure 13.4.
Planar drawings have applications in circuit layout and are helpful in displaying
graphical data such as program flow charts, organizational charts and scheduling
conflicts. For these applications, the goal is to draw the graph in the plane with as
few edge crossings as possible. (See the box on the following page for one such
example.)
13.2
Definitions of Planar Graphs
We took the idea of a planar drawing for granted in the previous section, but if
we’re going to prove things about planar graphs, we better have precise definitions.
Definition 13.2.1. A drawing of a graph assigns to each node a distinct point in
the plane and assigns to each edge a smooth curve in the plane whose endpoints
correspond to the nodes incident to the edge. The drawing is planar if none of the
“mcs” — 2017/6/5 — 19:42 — page 534 — #542
534
Chapter 13
Planar Graphs
Figure 13.1 Three dog houses and and three human houses. Is there a route from
each dog house to each human house so that no pair of routes cross each other?
“mcs” — 2017/6/5 — 19:42 — page 535 — #543
13.2. Definitions of Planar Graphs
535
Figure 13.2 Five quadrapi (4-armed creatures).
(a)
(b)
Figure 13.3 K3;3 (a) and K5 (b). Can you redraw these graphs so that no pairs
of edges cross?
“mcs” — 2017/6/5 — 19:42 — page 536 — #544
536
Chapter 13
Planar Graphs
u
v
u
v
(a)
(b)
Figure 13.4 Planar drawings of (a) K3;3 without hu—vi, and (b) K5 without
hu—vi.
Steve Wozniak and a Planar Circuit Design
When wires are arranged on a surface, like a circuit board or microchip, crossings require troublesome three-dimensional structures. When Steve Wozniak
designed the disk drive for the early Apple II computer, he struggled mightily to achieve a nearly planar design according to the following excerpt from
apple2history.org which in turn quotes Fire in the Valley by Freiberger
and Swaine:
For two weeks, he worked late each night to make a satisfactory design. When he was finished, he found that if he moved a connector
he could cut down on feedthroughs, making the board more reliable.
To make that move, however, he had to start over in his design. This
time it only took twenty hours. He then saw another feedthrough
that could be eliminated, and again started over on his design. “The
final design was generally recognized by computer engineers as brilliant and was by engineering aesthetics beautiful. Woz later said, ’It’s
something you can only do if you’re the engineer and the PC board
layout person yourself. That was an artistic layout. The board has
virtually no feedthroughs.’
“mcs” — 2017/6/5 — 19:42 — page 537 — #545
13.2. Definitions of Planar Graphs
537
curves cross themselves or other curves, namely, the only points that appear more
than once on any of the curves are the node points. A graph is planar when it has a
planar drawing.
Definition 13.2.1 is precise but depends on further concepts: “smooth planar
curves” and “points appearing more than once” on them. We haven’t defined these
concepts—we just showed the simple picture in Figure 13.4 and hoped you would
get the idea.
Pictures can be a great way to get a new idea across, but it is generally not a good
idea to use a picture to replace precise mathematics. Relying solely on pictures can
sometimes lead to disaster—or to bogus proofs, anyway. There is a long history of
bogus proofs about planar graphs based on misleading pictures.
The bad news is that to prove things about planar graphs using the planar drawings of Definition 13.2.1, we’d have to take a chapter-long excursion into continuous mathematics just to develop the needed concepts from plane geometry and
point-set topology. The good news is that there is another way to define planar
graphs that uses only discrete mathematics. In particular, we can define planar
graphs as a recursive data type. In order to understand how it works, we first need
to understand the concept of a face in a planar drawing.
13.2.1
Faces
The curves in a planar drawing divide up the plane into connected regions called
the continuous faces1 of the drawing. For example, the drawing in Figure 13.5 has
four continuous faces. Face IV, which extends off to infinity in all directions, is
called the outside face.
The vertices along the boundary of each continuous face in Figure 13.5 form a
cycle. For example, labeling the vertices as in Figure 13.6, the cycles for each of
the face boundaries can be described by the vertex sequences
abca
abda
bcdb
acda:
(13.1)
These four cycles correspond nicely to the four continuous faces in Figure 13.6—
so nicely, in fact, that we can identify each of the faces in Figure 13.6 by its cycle.
For example, the cycle abca identifies face III. The cycles in list 13.1 are called the
discrete faces of the graph in Figure 13.6. We use the term “discrete” since cycles
in a graph are a discrete data type—as opposed to a region in the plane, which is a
continuous data type.
1 Most
texts drop the adjective continuous from the definition of a face as a connected region. We
need the adjective to distinguish continuous faces from the discrete faces we’re about to define.
“mcs” — 2017/6/5 — 19:42 — page 538 — #546
538
Chapter 13
Planar Graphs
III
II
I
IV
Figure 13.5 A planar drawing with four continuous faces.
b
III
a
II
I
c
IV
d
Figure 13.6 The drawing with labeled vertices.
“mcs” — 2017/6/5 — 19:42 — page 539 — #547
13.2. Definitions of Planar Graphs
539
f
b
a
c
d
e
g
Figure 13.7 A planar drawing with a bridge.
Unfortunately, continuous faces in planar drawings are not always bounded by
cycles in the graph—things can get a little more complicated. For example, the
planar drawing in Figure 13.7 has what we will call a bridge, namely, a cut edge
hc—ei. The sequence of vertices along the boundary of the outer region of the
drawing is
abcefgecda:
This sequence defines a closed walk, but does not define a cycle since the walk has
two occurrences of the bridge hc—ei and each of its endpoints.
The planar drawing in Figure 13.8 illustrates another complication. This drawing
has what we will call a dongle, namely, the nodes v, x, y and w, and the edges
incident to them. The sequence of vertices along the boundary of the inner region
is
rst vxyxvwvt ur:
This sequence defines a closed walk, but once again does not define a cycle because it has two occurrences of every edge of the dongle—once “coming” and once
“going.”
It turns out that bridges and dongles are the only complications, at least for connected graphs. In particular, every continuous face in a planar drawing corresponds
to a closed walk in the graph. These closed walks will be called the discrete faces
of the drawing, and we’ll define them next.
13.2.2
A Recursive Definition for Planar Embeddings
The association between the continuous faces of a planar drawing and closed walks
provides the discrete data type we can use instead of continuous drawings. We’ll
define a planar embedding of connected graph to be the set of closed walks that are
its face boundaries. Since all we care about in a graph are the connections between
“mcs” — 2017/6/5 — 19:42 — page 540 — #548
540
Chapter 13
Planar Graphs
s
y
x
v
r
t
w
u
Figure 13.8 A planar drawing with a dongle.
vertices—not what a drawing of the graph actually looks like—planar embeddings
are exactly what we need.
The question is how to define planar embeddings without appealing to continuous drawings. There is a simple way to do this based on the idea that any continuous
drawing can drawn step by step:
either draw a new point somewhere in the plane to represent a vertex,
or draw a curve between two vertex points that have already been laid down,
making sure the new curve doesn’t cross any of the previously drawn curves.
A new curve won’t cross any other curves precisely when it stays within one
of the continuous faces. Alternatively, a new curve won’t have to cross any other
curves if it can go between the outer faces of two different drawings. So to be sure
it’s ok to draw a new curve, we just need to check that its endpoints are on the
boundary of the same face, or that its endpoints are on the outer faces of different
drawings. Of course drawing the new curve changes the faces slightly, so the face
boundaries will have to be updated once the new curve is drawn. This is the idea
behind the following recursive definition.
Definition 13.2.2. A planar embedding of a connected graph consists of a nonempty
set of closed walks of the graph called the discrete faces of the embedding. Planar
embeddings are defined recursively as follows:
Base case: If G is a graph consisting of a single vertex v then a planar embedding
of G has one discrete face, namely, the length zero closed walk v.
“mcs” — 2017/6/5 — 19:42 — page 541 — #549
13.2. Definitions of Planar Graphs
541
w
a
x
z
b
y
Figure 13.9 The “split a face” case: awxbyza splits into awxba and abyza.
Constructor case (split a face): Suppose G is a connected graph with a planar
embedding, and suppose a and b are distinct, nonadjacent vertices of G that occur
in some discrete face of the planar embedding. That is, is a closed walk of the
form
D ˛bˇ
where ˛ is a walk from a to b and ˇ is a walk from b to a. Then the graph obtained
by adding the edge ha—bi to the edges of G has a planar embedding with the same
discrete faces as G, except that face is replaced by the two discrete faces2
˛b hb—ai
and
ha—bi bˇ
(13.2)
as illustrated in Figure 13.9.3
Constructor case (add a bridge): Suppose G and H are connected graphs with
planar embeddings and disjoint sets of vertices. Let be a discrete face of the
embedding of G and suppose that begins and ends at vertex a.
Similarly, let ı be a discrete face of the embedding of H that begins and ends at
vertex b.
2 There is a minor exception to this definition of embedding in the special case when G is a line
graph beginning with a and ending with b. In this case the cycles into which splits are actually
the same. That’s because adding edge ha—bi creates a cycle that divides the plane into “inner” and
“outer” continuous faces that are both bordered by this cycle. In order to maintain the correspondence
between continuous faces and discrete faces in this case, we define the two discrete faces of the
embedding to be two “copies” of this same cycle.
3 Formally, merge is an operation on walks, not a walk and an edge, so in (13.2), we should have
used a walk .a ha—bi b/ instead of an edge ha—bi and written
˛b.b hb—ai a/
and
.a ha—bi b/bˇ
“mcs” — 2017/6/5 — 19:42 — page 542 — #550
542
Chapter 13
Planar Graphs
t
z
a
u
b
y
w
v
x
Figure 13.10 The “add a bridge” case.
Then the graph obtained by connecting G and H with a new edge ha—bi has a
planar embedding whose discrete faces are the union of the discrete faces of G and
H , except that faces and ı are replaced by one new face
b ha—bi bıb hb—ai :
This is illustrated in Figure 13.10, where the vertex sequences of the faces of G
and H are:
G W faxyza; axya; ayzag
H W fbt uvwb; bt vwb; t uvtg;
and after adding the bridge ha—bi, there is a single connected graph whose faces
have the vertex sequences
faxyzabt uvwba; axya; ayza; bt vwb; t uvtg:
A bridge is simply a cut edge, but in the context of planar embeddings, the
bridges are precisely the edges that occur twice on the same discrete face—as opposed to once on each of two faces. Dongles are trees made of bridges; we only
use dongles in illustrations, so there’s no need to define them more precisely.
13.2.3
Does It Work?
Yes! In general, a graph is planar because it has a planar drawing according to
Definition 13.2.1 if and only if each of its connected components has a planar embedding as specified in Definition 13.2.2. Of course we can’t prove this without an
excursion into exactly the kind of continuous math that we’re trying to avoid. But
now that the recursive definition of planar graphs is in place, we won’t ever need to
fall back on the continuous stuff. That’s the good news.
The bad news is that Definition 13.2.2 is a lot more technical than the intuitively
simple notion of a drawing whose edges don’t cross. In many cases it’s easier to
“mcs” — 2017/6/5 — 19:42 — page 543 — #551
13.2. Definitions of Planar Graphs
543
r
r
u
s s
u
t
t
Figure 13.11 Two illustrations of the same embedding.
stick to the idea of planar drawings and give proofs in those terms. For example,
erasing edges from a planar drawing will surely leave a planar drawing. On the
other hand, it’s not so obvious, though of course it is true, that you can delete an
edge from a planar embedding and still get a planar embedding (see Problem 13.9).
In the hands of experts, and perhaps in your hands too with a little more experience, proofs about planar graphs by appeal to drawings can be convincing and
reliable. But given the long history of mistakes in such proofs, it’s safer to work
from the precise definition of planar embedding. More generally, it’s also important
to see how the abstract properties of curved drawings in the plane can be modelled
successfully using a discrete data type.
13.2.4
Where Did the Outer Face Go?
Every planar drawing has an immediately-recognizable outer face —it’s the one
that goes to infinity in all directions. But where is the outer face in a planar embedding?
There isn’t one! That’s because there really isn’t any need to distinguish one face
from another. In fact, a planar embedding could be drawn with any given face on
the outside. An intuitive explanation of this is to think of drawing the embedding
on a sphere instead of the plane. Then any face can be made the outside face by
“puncturing” that face of the sphere, stretching the puncture hole to a circle around
the rest of the faces, and flattening the circular drawing onto the plane.
So pictures that show different “outside” boundaries may actually be illustrations of the same planar embedding. For example, the two embeddings shown in
Figure 13.11 are really the same—check it: they have the same boundary cycles.
This is what justifies the “add bridge” case in Definition 13.2.2: whatever face
is chosen in the embeddings of each of the disjoint planar graphs, we can draw
a bridge between them without needing to cross any other edges in the drawing,
because we can assume the bridge connects two “outer” faces.
“mcs” — 2017/6/5 — 19:42 — page 544 — #552
544
13.3
Chapter 13
Planar Graphs
Euler’s Formula
The value of the recursive definition is that it provides a powerful technique for
proving properties of planar graphs, namely, structural induction. For example,
we will now use Definition 13.2.2 and structural induction to establish one of the
most basic properties of a connected planar graph, namely, that the number of vertices and edges completely determines the number of faces in every possible planar
embedding of the graph.
Theorem 13.3.1 (Euler’s Formula). If a connected graph has a planar embedding,
then
v eCf D2
where v is the number of vertices, e is the number of edges and f is the number of
faces.
For example, in Figure 13.5, v D 4, e D 6 and f D 4. Sure enough, 4 6 C 4 D
2, as Euler’s Formula claims.
Proof. The proof is by structural induction on the definition of planar embeddings.
Let P .E/ be the proposition that v e C f D 2 for an embedding E.
Base case (E is the one-vertex planar embedding): By definition, v D 1, e D 0 and
f D 1, and 1 0 C 1 D 2, so P .E/ indeed holds.
Constructor case (split a face): Suppose G is a connected graph with a planar
embedding, and suppose a and b are distinct, nonadjacent vertices of G that appear
on some discrete face D a : : : b a of the planar embedding.
Then the graph obtained by adding the edge ha—bi to the edges of G has a
planar embedding with one more face and one more edge than G. So the quantity
v e C f will remain the same for both graphs, and since by structural induction
this quantity is 2 for G’s embedding, it’s also 2 for the embedding of G with the
added edge. So P holds for the constructed embedding.
Constructor case (add bridge): Suppose G and H are connected graphs with planar embeddings and disjoint sets of vertices. Then connecting these two graphs
with a bridge merges the two bridged faces into a single face, and leaves all other
faces unchanged. So the bridge operation yields a planar embedding of a connected
“mcs” — 2017/6/5 — 19:42 — page 545 — #553
13.4. Bounding the Number of Edges in a Planar Graph
545
graph with vG C vH vertices, eG C eH C 1 edges, and fG C fH
.vG C vH /
D .vG
.eG C eH C 1/ C .fG C fH
eG C fG / C .vH
D .2/ C .2/
2
eH C fH /
1 faces. Since
1/
2
(by structural induction hypothesis)
D 2;
v e C f remains equal to 2 for the constructed embedding. That is, P .E/ also
holds in this case.
This completes the proof of the constructor cases, and the theorem follows by
structural induction.
13.4
Bounding the Number of Edges in a Planar Graph
Like Euler’s formula, the following lemmas follow by structural induction directly
from Definition 13.2.2.
Lemma 13.4.1. In a planar embedding of a connected graph, each edge occurs
once in each of two different faces, or occurs exactly twice in one face.
Lemma 13.4.2. In a planar embedding of a connected graph with at least three
vertices, each face is of length at least three.
Combining Lemmas 13.4.1 and 13.4.2 with Euler’s Formula, we can now prove
that planar graphs have a limited number of edges:
Theorem 13.4.3. Suppose a connected planar graph has v 3 vertices and e
edges. Then
e 3v 6:
(13.3)
Proof. By definition, a connected graph is planar iff it has a planar embedding. So
suppose a connected graph with v vertices and e edges has a planar embedding
with f faces. By Lemma 13.4.1, every edge has exactly two occurrences in the
face boundaries. So the sum of the lengths of the face boundaries is exactly 2e.
Also by Lemma 13.4.2, when v 3, each face boundary is of length at least three,
so this sum is at least 3f . This implies that
3f 2e:
(13.4)
“mcs” — 2017/6/5 — 19:42 — page 546 — #554
546
Chapter 13
But f D e
Planar Graphs
v C 2 by Euler’s formula, and substituting into (13.4) gives
v C 2/ 2e
3.e
e
3v C 6 0
e 3v
13.5
6
Returning to K5 and K3;3
Finally we have a simple way to answer the quadrapi question at the beginning of
this chapter: the five quadrapi can’t all shake hands without crossing. The reason
is that we know the quadrupi question is the same as asking whether a complete
graph K5 is planar, and Theorem 13.4.3 has the immediate:
Corollary 13.5.1. K5 is not planar.
Proof. K5 is connected and has 5 vertices and 10 edges. But since 10 > 3 5
K5 does not satisfy the inequality (13.3) that holds in all planar graphs.
6,
We can also use Euler’s Formula to show that K3;3 is not planar. The proof is
similar to that of Theorem 13.3 except that we use the additional fact that K3;3 is a
bipartite graph.
Lemma 13.5.2. In a planar embedding of a connected bipartite graph with at least
3 vertices, each face has length at least 4.
Proof. By Lemma 13.4.2, every face of a planar embedding of the graph has length
at least 3. But by Lemma 12.6.2 and Theorem 12.8.3.3, a bipartite graph can’t have
odd length closed walks. Since the faces of a planar embedding are closed walks,
there can’t be any faces of length 3 in a bipartite embedding. So every face must
have length at least 4.
Theorem 13.5.3. Suppose a connected bipartite graph with v 3 vertices and e
edges is planar. Then
e 2v 4:
(13.5)
Proof. Lemma 13.5.2 implies that all the faces of an embedding of the graph have
length at least 4. Now arguing as in the proof of Theorem 13.4.3, we find that the
sum of the lengths of the face boundaries is exactly 2e and at least 4f . Hence,
4f 2e
(13.6)
“mcs” — 2017/6/5 — 19:42 — page 547 — #555
13.6. Coloring Planar Graphs
547
for any embedding of a planar bipartite graph. By Euler’s theorem, f D 2
Substituting 2 v C e for f in (13.6), we have
4.2
v C e.
v C e/ 2e;
which simplies to (13.5).
Corollary 13.5.4. K3;3 is not planar.
Proof. K3;3 is connected, bipartite and has 6 vertices and 9 edges. But since 9 >
2 6 4, K3;3 does not satisfy the inequality (13.3) that holds in all bipartite planar
graphs.
13.6
Coloring Planar Graphs
We’ve covered a lot of ground with planar graphs, but not nearly enough to prove
the famous 4-color theorem. But we can get awfully close. Indeed, we have done
almost enough work to prove that every planar graph can be colored using only 5
colors.
There are two familiar facts about planarity that we will need.
Lemma 13.6.1. Any subgraph of a planar graph is planar.
Lemma 13.6.2. Merging two adjacent vertices of a planar graph leaves another
planar graph.
Merging two adjacent vertices, n1 and n2 of a graph means deleting the two
vertices and then replacing them by a new “merged” vertex m adjacent to all the
vertices that were adjacent to either of n1 or n2 , as illustrated in Figure 13.12.
Many authors take Lemmas 13.6.1 and 13.6.2 for granted for continuous drawings of planar graphs described by Definition 13.2.1. With the recursive Definition 13.2.2 both Lemmas can actually be proved using structural induction (see
Problem 13.9).
We need only one more lemma:
Lemma 13.6.3. Every planar graph has a vertex of degree at most five.
Proof. Assuming to the contrary that every vertex of some planar graph had degree
at least 6, then the sum of the vertex degrees is at least 6v. But the sum of the
vertex degrees equals 2e by the Handshake Lemma 12.2.1, so we have e 3v
contradicting the fact that e 3v 6 < 3v by Theorem 13.4.3.
“mcs” — 2017/6/5 — 19:42 — page 548 — #556
548
Chapter 13
Planar Graphs
n1
n2
Figure 13.12
!
n1
n2
!
m
Merging adjacent vertices n1 and n2 into new vertex m.
Theorem 13.6.4. Every planar graph is five-colorable.
Proof. The proof will be by strong induction on the number v of vertices, with
induction hypothesis:
Every planar graph with v vertices is five-colorable.
Base cases (v 5): immediate.
Inductive case: Suppose G is a planar graph with v C 1 vertices. We will describe
a five-coloring of G.
First, choose a vertex g of G with degree at most 5; Lemma 13.6.3 guarantees
there will be such a vertex.
Case 1: (deg.g/ < 5): Deleting g from G leaves a graph H that is planar by
Lemma 13.6.1, and since H has v vertices, it is five-colorable by induction
hypothesis. Now define a five coloring of G as follows: use the five-coloring
of H for all the vertices besides g, and assign one of the five colors to g that
is not the same as the color assigned to any of its neighbors. Since there are
fewer than 5 neighbors, there will always be such a color available for g.
Case 2: (deg.g/ D 5): If the five neighbors of g in G were all adjacent to each
other, then these five vertices would form a nonplanar subgraph isomorphic
to K5 , contradicting Lemma 13.6.1 (since K5 is not planar). So there must
“mcs” — 2017/6/5 — 19:42 — page 549 — #557
13.7. Classifying Polyhedra
549
be two neighbors, n1 and n2 , of g that are not adjacent. Now merge n1 and g
into a new vertex, m. In this new graph, n2 is adjacent to m, and the graph is
planar by Lemma 13.6.2. So we can then merge m and n2 into a another new
vertex m0 , resulting in a new graph G 0 which by Lemma 13.6.2 is also planar.
Since G 0 has v 1 vertices, it is five-colorable by the induction hypothesis.
Now define a five coloring of G as follows: use the five-coloring of G 0 for
all the vertices besides g, n1 and n2 . Next assign the color of m0 in G 0 to
be the color of the neighbors n1 and n2 . Since n1 and n2 are not adjacent
in G, this defines a proper five-coloring of G except for vertex g. But since
these two neighbors of g have the same color, the neighbors of g have been
colored using fewer than five colors altogether. So complete the five-coloring
of G by assigning one of the five colors to g that is not the same as any of
the colors assigned to its neighbors.
13.7
Classifying Polyhedra
p
The Pythagoreans had two great mathematical secrets, the irrationality of 2 and
a geometric construct that we’re about to rediscover!
A polyhedron is a convex, three-dimensional region bounded by a finite number
of polygonal faces. If the faces are identical regular polygons and an equal number
of polygons meet at each corner, then the polyhedron is regular. Three examples
of regular polyhedra are shown in Figure 13.13: the tetrahedron, the cube, and the
octahedron.
We can determine how many more regular polyhedra there are by thinking about
planarity. Suppose we took any polyhedron and placed a sphere inside it. Then we
could project the polyhedron face boundaries onto the sphere, which would give
an image that was a planar graph embedded on the sphere, with the images of the
corners of the polyhedron corresponding to vertices of the graph. We’ve already
observed that embeddings on a sphere are the same as embeddings on the plane, so
Euler’s formula for planar graphs can help guide our search for regular polyhedra.
For example, planar embeddings of the three polyhedra in Figure 13.1 are shown
in Figure 13.14.
Let m be the number of faces that meet at each corner of a polyhedron, and let n
be the number of edges on each face. In the corresponding planar graph, there are
m edges incident to each of the v vertices. By the Handshake Lemma 12.2.1, we
“mcs” — 2017/6/5 — 19:42 — page 550 — #558
550
Chapter 13
Planar Graphs
(a)
Figure 13.13
(a)
Figure 13.14
dron (c).
(b)
(c)
The tetrahedron (a), cube (b), and octahedron (c).
v
(b)
(c)
Planar embeddings of the tetrahedron (a), cube (b), and octahe-
“mcs” — 2017/6/5 — 19:42 — page 551 — #559
13.7. Classifying Polyhedra
551
n m v e f polyhedron
3 3 4 6 4 tetrahedron
4 3 8 12 6 cube
3 4 6 12 8 octahedron
3 5 12 30 20 icosahedron
5 3 20 30 12 dodecahedron
Figure 13.15
The only possible regular polyhedra.
know:
mv D 2e:
Also, each face is bounded by n edges. Since each edge is on the boundary of two
faces, we have:
nf D 2e
Solving for v and f in these equations and then substituting into Euler’s formula
gives:
2e
2e
eC
D2
m
n
which simplifies to
1
1
1 1
C D C
(13.7)
m n
e
2
Equation 13.7 places strong restrictions on the structure of a polyhedron. Every
nondegenerate polygon has at least 3 sides, so n 3. And at least 3 polygons
must meet to form a corner, so m 3. On the other hand, if either n or m were
6 or more, then the left side of the equation could be at most 1=3 C 1=6 D 1=2,
which is less than the right side. Checking the finitely-many cases that remain turns
up only five solutions, as shown in Figure 13.15. For each valid combination of n
and m, we can compute the associated number of vertices v, edges e, and faces f .
And polyhedra with these properties do actually exist. The largest polyhedron, the
dodecahedron, was the other great mathematical secret of the Pythagorean sect.
The 5 polyhedra in Figure 13.15 are the only possible regular polyhedra. So if
you want to put more than 20 geocentric satellites in orbit so that they uniformly
blanket the globe—tough luck!
“mcs” — 2017/6/5 — 19:42 — page 552 — #560
552
13.8
Chapter 13
Planar Graphs
Another Characterization for Planar Graphs
We did not pick K5 and K3;3 as examples because of their application to dog
houses or quadrapi shaking hands. We really picked them because they provide
another, famous, discrete characterizarion of planar graphs:
Theorem 13.8.1 (Kuratowski). A graph is not planar if and only if it contains K5
or K3;3 as a minor.
Definition 13.8.2. A minor of a graph G is a graph that can be obtained by repeatedly4 deleting vertices, deleting edges, and merging adjacent vertices of G.
For example, Figure 13.16 illustrates why C3 is a minor of the graph in Figure 13.16(a). In fact C3 is a minor of a connected graph G if and only if G is not a
tree.
The known proofs of Kuratowski’s Theorem 13.8.1 are a little too long to include
in an introductory text, so we won’t give one.
Problems for Section 13.2
Practice Problems
Problem 13.1.
What are the discrete faces of the following two graphs?
Write each cycle as a sequence of letters without spaces, starting with the alphabetically earliest letter in the clockwise direction, for example “adbfa.” Separate
the sequences with spaces.
(a)
f
b
a
c
d
e
g
(b)
4 The
three operations can each be performed any number of times in any order.
“mcs” — 2017/6/5 — 19:42 — page 553 — #561
13.8. Another Characterization for Planar Graphs
553
v2
e1
v1
(a)
(b)
(c)
(e)
(f)
v3
e2
(d)
Figure 13.16 One method by which the graph in (a) can be reduced to C3 (f),
thereby showing that C3 is a minor of the graph. The steps are: merging the nodes
incident to e1 (b), deleting v1 and all edges incident to it (c), deleting v2 (d), deleting e2 , and deleting v3 (f).
“mcs” — 2017/6/5 — 19:42 — page 554 — #562
554
Chapter 13
Planar Graphs
s
y
x
v
r
t
w
u
Problems for Section 13.8
Exam Problems
Problem 13.2.
h
l
c
d
j
g
b
e
i
n o
m
k
a
g1
f
g2
g3
(a) Describe an isomorphism between graphs G1 and G2 , and another isomorphism between G2 and G3 .
(b) Why does part .a/ imply that there is an isomorphism between graphs G1 and
G3 ?
Let G and H be planar graphs. An embedding EG of G is isomorphic to an embedding EH of H iff there is an isomorphism from G to H that also maps each
face of EG to a face of EH .
(c) One of the embeddings pictured above is not isomorphic to either of the others.
Which one? Briefly explain why.
(d) Explain why all embeddings of two isomorphic planar graphs must have the
“mcs” — 2017/6/5 — 19:42 — page 555 — #563
13.8. Another Characterization for Planar Graphs
555
same number of faces.
Problem 13.3. (a) Give an example of a planar graph with two planar embeddings,
where the first embedding has a face whose length is not equal to the length of any
face in the secoind embedding. Draw the two embeddings to demonstrate this.
(b) Define the length of a planar embedding E to be the sum of the lengths of
the faces of E. Prove that all embeddings of the same planar graph have the same
length.
Problem 13.4.
Definition 13.2.2 of planar graph embeddings applied only to connected planar
graphs. The definition can be extended to planar graphs that are not necessarily
connected by adding the following additional constructor case to the definition:
Constructor Case: (collect disjoint graphs) Suppose E1 and E2 are planar
embeddings with no vertices in common. Then E1 [ E2 is a planar embedding.
Euler’s Planar Graph Theorem now generalizes to unconnected graphs as follows: if a planar embedding E has v vertices, e edges, f faces and c connected
components, then
v e C f 2c D 0:
(13.8)
This can be proved by structural induction on the definition of planar embedding.
(a) State and prove the base case of the structural induction.
(b) Let vi ; ei ; fi ; and ci be the number of vertices, edges, faces, and connected
components in embedding Ei and let v; e; f; c be the numbers for the embedding
from the (collect disjoint graphs) constructor case. Express v; e; f; c in terms of
vi ; ei ; fi ; ci .
(c) Prove the (collect disjoint graphs) case of the structural induction.
Problem 13.5. (a) A simple graph has 8 vertices and 24 edges. What is the average
degree per vertex?
(b) A connected planar simple graph has 5 more edges than it has vertices. How
many faces does it have?
“mcs” — 2017/6/5 — 19:42 — page 556 — #564
556
Chapter 13
Planar Graphs
(c) A connected simple graph has one more vertex than it has edges. Explain why
it is a planar graph.
(d) How many faces does a planar graph from part c have?
(e) How many distinct isomorphisms are there between the graph given in Figure 13.17 and itself? (Include the identity isomorphism.)
a
d
e
b
f
c
Figure 13.17
Class Problems
x
Problem 13.6.
Figure 13.18 shows four different pictures of planar graphs.
(a) For each picture, describe its discrete faces (closed walks that define the region
borders).
(b) Which of the pictured graphs are isomorphic? Which pictures represent the
same planar embedding?—that is, they have the same discrete faces.
(c) Describe a way to construct the embedding in Figure 4 according to the recursive Definition 13.2.2 of planar embedding. For each application of a constructor
rule, be sure to indicate the faces (cycles) to which the rule was applied and the
cycles which result from the application.
Problem 13.7.
Prove the following assertions by structural induction on the definition of planar
embedding.
(a) In a planar embedding of a graph, each edge occurs exactly twice in the faces
of the embedding.
“mcs” — 2017/6/5 — 19:42 — page 557 — #565
13.8. Another Characterization for Planar Graphs
b
557
b
c
c
d
a
d
a
figure 2
figure 1
b
b
c
c
d
a
d
a
e
e
figure 3
figure 4
Figure 13.18
(b) In a planar embedding of a connected graph with at least three vertices, each
face is of length at least three.
Homework Problems
Problem 13.8.
A simple graph is triangle-free when it has no cycle of length three.
(a) Prove for any connected triangle-free planar graph with v > 2 vertices and e
edges,
e 2v 4:
(13.9)
(b) Show that any connected triangle-free planar graph has at least one vertex of
degree three or less.
“mcs” — 2017/6/5 — 19:42 — page 558 — #566
558
Chapter 13
Planar Graphs
(c) Prove that any connected triangle-free planar graph is 4-colorable.
Problem 13.9. (a) Prove
Lemma (Switch Edges). Suppose that, starting from some embeddings of planar
graphs with disjoint sets of vertices, it is possible by two successive applications of
constructor operations to add edges e and then f to obtain a planar embedding F.
Then starting from the same embeddings, it is also possible to obtain F by adding
f and then e with two successive applications of constructor operations.
Hint: There are four cases to analyze, depending on which two constructor operations are applied to add e and then f . Structural induction is not needed.
(b) Prove
Corollary (Permute Edges). Suppose that, starting from some embeddings of planar graphs with disjoint sets of vertices, it is possible to add a sequence of edges
e0 ; e1 ; : : : ; en by successive applications of constructor operations to obtain a planar embedding F. Then starting from the same embeddings, it is also possible
to obtain F by applications of constructor operations that successively add any
permutation5 of the edges e0 ; e1 ; : : : ; en .
Hint: By induction on the number of switches of adjacent elements needed to convert the sequence 0,1,. . . ,n into a permutation .0/; .1/; : : : ; .n/.
(c) Prove
Corollary (Delete Edge). Deleting an edge from a planar graph leaves a planar
graph.
(d) Conclude that any subgraph of a planar graph is planar.
5 If
W f0; 1; : : : ; ng ! f0; 1; : : : ; ng is a bijection, then the sequence e.0/ ; e.1/ ; : : : ; e.n/ is
called a permutation of the sequence e0 ; e1 ; : : : ; en .
“mcs” — 2017/6/5 — 19:42 — page 559 — #567
III
Counting
“mcs” — 2017/6/5 — 19:42 — page 560 — #568
“mcs” — 2017/6/5 — 19:42 — page 561 — #569
Introduction
Counting seems easy enough: 1, 2, 3, 4, etc. This direct approach works well
for counting simple things—like your toes—and may be the only approach for extremely complicated things with no identifiable structure. However, subtler methods can help you count many things in the vast middle ground, such as:
The number of different ways to select a dozen doughnuts when there are
five varieties available.
The number of 16-bit numbers with exactly 4 ones.
Perhaps surprisingly, but certainly not coincidentally, these two numbers are the
same: 1820.
Counting is useful in computer science for several reasons:
Determining the time and storage required to solve a computational problem—
a central objective in computer science—often comes down to solving a
counting problem.
Password and encryption security counts on having a very large set of possible passwords and encryption keys.
Counting is the basis of probability theory, which plays a central role in all
sciences, including computer science.
We begin our study of counting in Chapter 14 with a collection of rules and
methods for finding
P closed-formQexpressions for commonly-occurring sums and
products such as niD0 x i and niD1 i . We also introduce asymptotic notations
such as , O and ‚ that are commonly used in computer science to express how a
quantity such as the running time of a program grows with the size of the input.
“mcs” — 2017/6/5 — 19:42 — page 562 — #570
562
Part III
Counting
Chapter 15 describes the most basic rules for determining the cardinality of a set.
These rules are actually theorems, but our focus here will be less on their proofs
than on teaching their use in simple counting as a practical skill, like integration.
But counting can be tricky, and people make counting mistakes all the time,
so a crucial part of counting skill is being able to verify a counting argument.
Sometimes this can be done simply by finding an alternative way to count and
then comparing answers—they better agree. But most elementary counting arguments reduce to finding a bijection between objects to be counted and easy-to-count
sequences. The chapter shows how explicitly defining these bijections—and verifying that they are bijections—is another useful way to verify counting arguments.
The material in Chapter 15 is simple yet powerful, and it provides a great tool set
for use in your future career.
Finally, Chapter 16 introduces generating functions which allow many counting
problems to be solved by simple algebraic formula simplification.
“mcs” — 2017/6/5 — 19:42 — page 563 — #571
14
Sums and Asymptotics
Sums and products arise regularly in the analysis of algorithms, financial applications, physical problems, and probabilistic systems. For example, according to
Theorem 2.2.1,
n.n C 1/
1 C 2 C 3 C C n D
:
(14.1)
2
Of course, the left-hand sum could be expressed concisely as a subscripted summation
n
X
i
i D1
but the right-hand expression n.n C 1/=2 is not only concise but also easier to
evaluate. Furthermore, it more clearly reveals properties such as the growth rate
of the sum. Expressions like n.n C 1/=2 that do not make use of subscripted
summations or products—or those handy but sometimes troublesome sequences
of three dots—are called closed forms.
Another example is the closed form for a geometric sum
1 C x C x2 C x3 C C xn D
1
x nC1
1 x
(14.2)
given in Problem 5.4. The sum as described on the left-hand side of (14.2) involves
n additions and 1 C 2 C C .n 1/ D .n 1/n=2 multiplications, but its closed
form on the right-hand side can be evaluated using fast exponentiation with at most
2 log n multiplications, a division, and a couple of subtractions. Also, the closed
form makes the growth and limiting behavior of the sum much more apparent.
Equations (14.1) and (14.2) were easy to verify by induction, but, as is often the
case, the proofs by induction gave no hint about how these formulas were found in
the first place. Finding them is part math and part art, which we’ll start examining
in this chapter.
Our first motivating example will be the value of a financial instrument known as
an annuity. This value will be a large and nasty-looking sum. We will then describe
several methods for finding closed forms for several sorts of sums, including those
for annuities. In some cases, a closed form for a sum may not exist, and so we
will provide a general method for finding closed forms for good upper and lower
bounds on the sum.
The methods we develop for sums will also work for products, since any product
can be converted into a sum by taking its logarithm. For instance, later in the
“mcs” — 2017/6/5 — 19:42 — page 564 — #572
564
Chapter 14
Sums and Asymptotics
chapter we will use this approach to find a good closed-form approximation to the
factorial function
nŠ WWD 1 2 3 n:
We conclude the chapter with a discussion of asymptotic notation, especially
“Big Oh” notation. Asymptotic notation is often used to bound the error terms
when there is no exact closed form expression for a sum or product. It also provides
a convenient way to express the growth rate or order of magnitude of a sum or
product.
14.1
The Value of an Annuity
Would you prefer a million dollars today or $50,000 a year for the rest of your life?
On the one hand, instant gratification is nice. On the other hand, the total dollars
received at $50K per year is much larger if you live long enough.
Formally, this is a question about the value of an annuity. An annuity is a financial instrument that pays out a fixed amount of money at the beginning of every year
for some specified number of years. In particular, an n-year, m-payment annuity
pays m dollars at the start of each year for n years. In some cases, n is finite, but
not always. Examples include lottery payouts, student loans, and home mortgages.
There are even firms on Wall Street that specialize in trading annuities.1
A key question is, “What is an annuity worth?” For example, lotteries often pay
out jackpots over many years. Intuitively, $50,000 a year for 20 years ought to be
worth less than a million dollars right now. If you had all the cash right away, you
could invest it and begin collecting interest. But what if the choice were between
$50,000 a year for 20 years and a half million dollars today? Suddenly, it’s not
clear which option is better.
14.1.1
The Future Value of Money
In order to answer such questions, we need to know what a dollar paid out in the
future is worth today. To model this, let’s assume that money can be invested at a
fixed annual interest rate p. We’ll assume an 8% rate2 for the rest of the discussion,
so p D 0:08.
1 Such
trading ultimately led to the subprime mortgage disaster in 2008–2009. We’ll talk more
about that in a later chapter.
2 U.S. interest rates have dropped steadily for several years, and ordinary bank deposits now earn
around 1.0%. But just a few years ago the rate was 8%; this rate makes some of our examples a little
more dramatic. The rate has been as high as 17% in the past thirty years.
“mcs” — 2017/6/5 — 19:42 — page 565 — #573
14.1. The Value of an Annuity
565
Here is why the interest rate p matters. Ten dollars invested today at interest rate
p will become .1 C p/ 10 D 10:80 dollars in a year, .1 C p/2 10 11:66 dollars
in two years, and so forth. Looked at another way, ten dollars paid out a year from
now is only really worth 1=.1 C p/ 10 9:26 dollars today, because if we had the
$9.26 today, we could invest it and would have $10.00 in a year anyway. Therefore,
p determines the value of money paid out in the future.
So for an n-year, m-payment annuity, the first payment of m dollars is truly worth
m dollars. But the second payment a year later is worth only m=.1 C p/ dollars.
Similarly, the third payment is worth m=.1 C p/2 , and the n-th payment is worth
only m=.1 C p/n 1 . The total value V of the annuity is equal to the sum of the
payment values. This gives:
n
X
m
.1 C p/i 1
i D1
n
X1 1 j
Dm
1Cp
V D
(substitute j D i
1)
j D0
Dm
n
X1
xj
(substitute x D 1=.1 C p/):
(14.3)
j D0
The goal of the preceding substitutions was to get the summation into the form
of a simple geometric sum. This leads us to an explanation of a way you could have
discovered the closed form (14.2) in the first place using the Perturbation Method.
14.1.2
The Perturbation Method
Given a sum that has a nice structure, it is often useful to “perturb” the sum so that
we can somehow combine the sum with the perturbation to get something much
simpler. For example, suppose
S D 1 C x C x2 C C xn:
An example of a perturbation would be
xS D x C x 2 C C x nC1 :
The difference between S and xS is not so great, and so if we were to subtract xS
from S , there would be massive cancellation:
S D 1 C x C x2 C x3 C C xn
xS D
x x2 x3 xn
x nC1 :
“mcs” — 2017/6/5 — 19:42 — page 566 — #574
566
Chapter 14
Sums and Asymptotics
The result of the subtraction is
xS D 1
S
x nC1 :
Solving for S gives the desired closed-form expression in equation 14.2, namely,
SD
1
x nC1
:
1 x
We’ll see more examples of this method when we introduce generating functions
in Chapter 16.
14.1.3
A Closed Form for the Annuity Value
Using equation 14.2, we can derive a simple formula for V , the value of an annuity
that pays m dollars at the start of each year for n years.
1 xn
(by equations 14.3 and 14.2)
(14.4)
V Dm
1 x
!
1 C p .1=.1 C p//n 1
Dm
(substituting x D 1=.1 C p/): (14.5)
p
Equation 14.5 is much easier to use than a summation with dozens of terms. For
example, what is the real value of a winning lottery ticket that pays $50,000 per
year for 20 years? Plugging in m D $50,000, n D 20 and p D 0:08 gives
V $530,180. So because payments are deferred, the million dollar lottery is
really only worth about a half million dollars! This is a good trick for the lottery
advertisers.
14.1.4
Infinite Geometric Series
We began this chapter by asking whether you would prefer a million dollars today
or $50,000 a year for the rest of your life. Of course, this depends on how long
you live, so optimistically assume that the second option is to receive $50,000 a
year forever. This sounds like infinite money! But we can compute the value of an
annuity with an infinite number of payments by taking the limit of our geometric
sum in equation 14.2 as n tends to infinity.
Theorem 14.1.1. If jxj < 1, then
1
X
i D0
xi D
1
1
x
:
“mcs” — 2017/6/5 — 19:42 — page 567 — #575
14.1. The Value of an Annuity
567
Proof.
1
X
x i WWD lim
n
X
n!1
i D0
iD0
1
D lim
n!1
D
1
1
x
xi
x nC1
1 x
(by equation 14.2)
:
The final line follows from the fact that limn!1 x nC1 D 0 when jxj < 1.
In our annuity problem x D 1=.1 C p/ < 1, so Theorem 14.1.1 applies, and we
get
V Dm
1
X
xj
(by equation 14.3)
j D0
1
1 x
1Cp
Dm
p
Dm
(by Theorem 14.1.1)
.x D 1=.1 C p//:
Plugging in m D $50,000 and p D 0:08, we see that the value V is only $675,000.
It seems amazing that a million dollars today is worth much more than $50,000
paid every year for eternity! But on closer inspection, if we had a million dollars
today in the bank earning 8% interest, we could take out and spend $80,000 a year,
forever. So as it turns out, this answer really isn’t so amazing after all.
14.1.5
Examples
Equation 14.2 and Theorem 14.1.1 are incredibly useful in computer science.
Here are some other common sums that can be put into closed form using equa-
“mcs” — 2017/6/5 — 19:42 — page 568 — #576
568
Chapter 14
Sums and Asymptotics
tion 14.2 and Theorem 14.1.1:
1 i
X
1
1
D2
2
1 .1=2/
i D0
!
!
1 X
1
10
1 i
0:99999 D 0:9
D 0:9
D 0:9
D1
10
1 1=10
9
1 C 1=2 C 1=4 C D
D
(14.6)
(14.7)
i D0
1
1=2 C 1=4
1 X
1 i
D
D
2
1
i D0
1 C 2 C 4 C C 2n
1
D
n
X1
2i D
i D0
1 C 3 C 9 C C 3n
1
D
n
X1
3i D
i D0
1
. 1=2/
D
1 2n
D 2n
1 2
1 3n
3n 1
D
1 3
2
2
3
(14.8)
1
(14.9)
(14.10)
If the terms in a geometric sum grow smaller, as in equation 14.6, then the sum is
said to be geometrically decreasing. If the terms in a geometric sum grow progressively larger, as in equations 14.9 and 14.10, then the sum is said to be geometrically
increasing. In either case, the sum is usually approximately equal to the term in the
sum with the greatest absolute value. For example, in equations 14.6 and 14.8, the
largest term is equal to 1 and the sums are 2 and 2/3, both relatively close to 1. In
equation 14.9, the sum is about twice the largest term. In equation 14.10, the largest
term is 3n 1 and the sum is .3n 1/=2, which is only about a factor of 1:5 greater.
You can see why this rule of thumb works by looking carefully at equation 14.2
and Theorem 14.1.1.
14.1.6
Variations of Geometric Sums
We now know all about geometric sums—if you have one, life is easy. But in
practice one often encounters
P i sums that cannot be transformed by simple variable
substitutions to the form x .
A non-obvious but useful way to obtain new summation formulas from old ones
is by differentiating or integrating with respect to x. As an example, consider the
following sum:
n
X1
ix i D x C 2x 2 C 3x 3 C C .n
1/x n
1
i D1
This is not a geometric sum. The ratio between successive terms is not fixed, and
so our formula for the sum of a geometric sum cannot be directly applied. But
“mcs” — 2017/6/5 — 19:42 — page 569 — #577
14.1. The Value of an Annuity
569
differentiating equation 14.2 leads to:
!
n 1
d X i
d 1 xn
:
x D
dx
dx 1 x
(14.11)
i D0
The left-hand side of equation 14.11 is simply
n
X1
i D0
n
X1
d i
.x / D
ix i
dx
1
:
i D0
The right-hand side of equation 14.11 is
nx n
1 .1
x/ . 1/.1
.1 x/2
xn/
nx n
D
D
1
1
C nx n C 1 x n
.1 x/2
n
nx 1 C .n 1/x n
:
.1 x/2
Hence, equation 14.11 means that
n
X1
ix i
1
1
D
i D0
nx n 1 C .n
.1 x/2
1/x n
:
Incidentally, Problem 14.2 shows how the perturbation method could also be applied to derive this formula.
Often, differentiating or integrating messes up the exponent
x in every term.
P of
i
1
In this case, we now have
P a formula for a sum of the form ix , but we want a
formula for the series ix i . The solution is simple: multiply by x. This gives:
n
X1
ix i D
x
i D1
nx n C .n 1/x nC1
.1 x/2
(14.12)
and we have the desired closed-form expression for our sum. It seems a little complicated, but it’s easier to work with than the sum.
Notice that if jxj < 1, then this series converges to a finite value even if there
are infinitely many terms. Taking the limit of equation 14.12 as n tends to infinity
gives the following theorem:
Theorem 14.1.2. If jxj < 1, then
1
X
i D1
ix i D
x
.1
x/2
:
(14.13)
“mcs” — 2017/6/5 — 19:42 — page 570 — #578
570
Chapter 14
Sums and Asymptotics
As a consequence, suppose that there is an annuity that pays i m dollars at the
end of each year i , forever. For example, if m D $50,000, then the payouts are
$50,000 and then $100,000 and then $150,000 and so on. It is hard to believe that
the value of this annuity is finite! But we can use Theorem 14.1.2 to compute the
value:
V D
1
X
i D1
Dm
Dm
im
.1 C p/i
1=.1 C p/
.1
1 2
1Cp /
1Cp
:
p2
The second line follows by an application of Theorem 14.1.2. The third line is
obtained by multiplying the numerator and denominator by .1 C p/2 .
For example, if m D $50,000, and p D 0:08 as usual, then the value of the
annuity is V D $8,437,500. Even though the payments increase every year, the increase is only additive with time; by contrast, dollars paid out in the future decrease
in value exponentially with time. The geometric decrease swamps out the additive
increase. Payments in the distant future are almost worthless, so the value of the
annuity is finite.
The important thing to remember is the trick of taking the derivative (or integral)
of a summation formula. Of course, this technique requires one to compute nasty
derivatives correctly, but this is at least theoretically possible!
14.2
Sums of Powers
In Chapter 5, we verified the formula (14.1), but the source of this formula is still
a mystery. Sure, we can prove that it’s true by using well ordering or induction,
but where did the expression on the right come from in the first place? Even more
inexplicable is the closed form expression for the sum of consecutive squares:
n
X
i D1
i2 D
.2n C 1/.n C 1/n
:
6
(14.14)
It turns out that there is a way to derive these expressions, but before we explain
it, we thought it would be fun—OK, our definition of “fun” may be different than
“mcs” — 2017/6/5 — 19:42 — page 571 — #579
14.2. Sums of Powers
571
yours—to show you how Gauss is supposed to have proved equation 14.1 when he
was a young boy.
Gauss’s idea is related to the perturbation method we used in Section 14.1.2. Let
SD
n
X
i:
i D1
Then we can write the sum in two orders:
S D 1 C 2 C : : : C .n 1/ C n;
S D n C .n 1/ C : : : C 2 C 1:
Adding these two equations gives
2S D .n C 1/ C .n C 1/ C C .n C 1/ C .n C 1/
D n.n C 1/:
Hence,
n.n C 1/
:
2
Not bad for a young child—Gauss showed some potential. . . .
Unfortunately, the same trick does not work for summing consecutive squares.
However, we can observe that the result might be a third-degree polynomial in n,
since the sum contains n terms that average out to a value that grows quadratically
in n. So we might guess that
SD
n
X
i 2 D an3 C bn2 C cn C d:
i D1
If our guess is correct, then we can determine the parameters a, b, c and d by
plugging in a few values for n. Each such value gives a linear equation in a, b,
c and d . If we plug in enough values, we may get a linear system with a unique
solution. Applying this method to our example gives:
nD0
implies 0 D d
nD1
implies 1 D a C b C c C d
nD2
implies 5 D 8a C 4b C 2c C d
nD3
implies 14 D 27a C 9b C 3c C d:
Solving this system gives the solution a D 1=3, b D 1=2, c D 1=6, d D 0.
Therefore, if our initial guess at the form of the solution was correct, then the
summation is equal to n3 =3 C n2 =2 C n=6, which matches equation 14.14.
“mcs” — 2017/6/5 — 19:42 — page 572 — #580
572
Chapter 14
Sums and Asymptotics
The point is that if the desired formula turns out to be a polynomial, then once
you get an estimate of the degree of the polynomial, all the coefficients of the
polynomial can be found automatically.
Be careful! This method lets you discover formulas, but it doesn’t guarantee
they are right! After obtaining a formula by this method, it’s important to go back
and prove it by induction or some other method. If the initial guess at the solution
was not of the right form, then the resulting formula will be completely wrong! A
later chapter will describe a method based on generating functions that does not
require any guessing at all.
14.3
Approximating Sums
Unfortunately, it is not always possible to find a closed-form expression for a sum.
For example, no closed form is known for
SD
n
X
p
i:
i D1
In such cases, we need to resort to approximations for S if we want to have a
closed form. The good news is that there is a general method to find closed-form
upper and lower bounds that works well for many sums. Even better, the method
is simple and easy to remember. It works by replacing the sum by an integral and
then adding either the first or last term in the sum.
Definition 14.3.1. A function f W RC ! RC is strictly increasing when
x < y IMPLIES f .x/ < f .y/;
and it is weakly increasing3 when
x < y IMPLIES f .x/ f .y/:
Similarly, f is strictly decreasing when
x < y IMPLIES f .x/ > f .y/;
and it is weakly decreasing4 when
x < y IMPLIES f .x/ f .y/:
3 Weakly increasing functions are usually called nondecreasing functions. We will avoid this
terminology to prevent confusion between being a nondecreasing function and the much weaker
property of not being a decreasing function.
4 Weakly decreasing functions are usually called nonincreasing.
“mcs” — 2017/6/5 — 19:42 — page 573 — #581
14.3. Approximating Sums
573
f.n/

f.n�1/
f.3/
f.2/
f.1/
0
Figure 14.1
P
n
i D1 f .i /.
1
2

3
n�2
n�1
n
The area of the i th rectangle is f .i /. The shaded region has area
p
For example, 2x and x are strictly increasing functions, while maxfx; 2g and
dxe are weakly increasing functions. The functions 1=x and 2 x are strictly decreasing, while minf1=x; 1=2g and b1=xc are weakly decreasing.
Theorem 14.3.2. Let f W RC ! RC be a weakly increasing function. Define
S WWD
n
X
f .i /
(14.15)
i D1
and
Z
I WWD
n
f .x/ dx:
1
Then
I C f .1/ S I C f .n/:
(14.16)
Similarly, if f is weakly decreasing, then
I C f .n/ S I C f .1/:
Proof. Suppose f W RC ! RC is weakly increasing. The value of the sum S
in (14.15) is the sum of the areas of n unit-width rectangles of heights f .1/; f .2/; : : : ; f .n/.
This area of these rectangles is shown shaded in Figure 14.1.
The value of
Z n
I D
f .x/ dx
1
is the shaded area under the curve of f .x/ from 1 to n shown in Figure 14.2.
“mcs” — 2017/6/5 — 19:42 — page 574 — #582
Chapter 14
Sums and Asymptotics
f.n/
f.n�1/
f.x/

574
f.3/
f.2/
f.1/
0
1
2

3
n�2
n�1
n
Figure R14.2 The shaded area under the curve of f .x/ from 1 to n (shown in bold)
n
is I D 1 f .x/ dx.
Comparing the shaded regions in Figures 14.1 and 14.2 shows that S is at least
I plus the area of the leftmost rectangle. Hence,
S I C f .1/
(14.17)
This is the lower bound for S given in (14.16).
To derive the upper bound for S given in (14.16), we shift the curve of f .x/
from 1 to n one unit to the left as shown in Figure 14.3.
Comparing the shaded regions in Figures 14.1 and 14.3 shows that S is at most
I plus the area of the rightmost rectangle. That is,
S I C f .n/;
which is the upper bound for S given in (14.16).
The very similar argument for the weakly decreasing case is left to Problem 14.10.
Theorem 14.3.2 provides good bounds for most sums. At worst, the bounds will
be off by the largest term in the sum. For example, we can use Theorem 14.3.2 to
bound the sum
n
X
p
SD
i
i D1
as follows.
“mcs” — 2017/6/5 — 19:42 — page 575 — #583
14.3. Approximating Sums
575
f.n/
f.xC1/

f.n�1/
f.3/
f.2/
f.1/
0
1
2

3
n�2
n�1
n
Figure 14.3 This curve is the same as the curve in Figure 14.2 shifted left by 1.
We begin by computing
Z
I D
np
x dx
ˇn
x 3=2 ˇˇ
D
ˇ
3=2 ˇ
1
1
2
D .n3=2
3
1/:
We then apply Theorem 14.3.2 to conclude that
2 3=2
.n
3
and thus that
1/ C 1 S 2 3=2
.n
3
1/ C
p
2 3=2 1
2
n
C S n3=2 C n
3
3
3
p
n
2
:
3
In other words, the sum is very close to 23 n3=2 . We’ll define several ways that one
thing can be “very close to” something else at the end of this chapter.
As a first application of Theorem 14.3.2, we explain in the next section how it
helps in resolving a classic paradox in structural engineering.
“mcs” — 2017/6/5 — 19:42 — page 576 — #584
576
14.4
Chapter 14
Sums and Asymptotics
Hanging Out Over the Edge
Suppose you have a bunch of books and you want to stack them up, one on top
of another in some off-center way, so the top book sticks out past books below it
without falling over. If you moved the stack to the edge of a table, how far past
the edge of the table do you think you could get the top book to go? Could the top
book stick out completely beyond the edge of table? You’re not supposed to use
glue or any other support to hold the stack in place.
Most people’s first response to the Book Stacking Problem—sometimes also
their second and third responses—is “No, the top book will never get completely
past the edge of the table.” But in fact, you can get the top book to stick out as far
as you want: one booklength, two booklengths, any number of booklengths!
14.4.1
Formalizing the Problem
We’ll approach this problem recursively. How far past the end of the table can we
get one book to stick out? It won’t tip as long as its center of mass is over the table,
so we can get it to stick out half its length, as shown in Figure 14.4.
center of mass
of book
1
2
table
Figure 14.4 One book can overhang half a book length.
Now suppose we have a stack of books that will not tip over if the bottom book
rests on the table—call that a stable stack. Let’s define the overhang of a stable
stack to be the horizontal distance from the center of mass of the stack to the furthest
edge of the top book. So the overhang is purely a property of the stack, regardless
of its placement on the table. If we place the center of mass of the stable stack at
the edge of the table as in Figure 14.5, the overhang is how far we can get the top
“mcs” — 2017/6/5 — 19:42 — page 577 — #585
14.4. Hanging Out Over the Edge
center of mass
of the whole stack
577
overhang
table
Figure 14.5 Overhanging the edge of the table.
book in the stack to stick out past the edge.
In general, a stack of n books will be stable if and only if the center of mass of
the top i books sits over the .i C 1/st book for i D 1, 2, . . . , n 1.
So we want a formula for the maximum possible overhang Bn achievable with a
stable stack of n books.
We’ve already observed that the overhang of one book is 1/2 a book length. That
is,
1
B1 D :
2
Now suppose we have a stable stack of n C 1 books with maximum overhang.
If the overhang of the n books on top of the bottom book was not maximum, we
could get a book to stick out further by replacing the top stack with a stack of n
books with larger overhang. So the maximum overhang BnC1 of a stack of n C 1
books is obtained by placing a maximum overhang stable stack of n books on top
of the bottom book. And we get the biggest overhang for the stack of n C 1 books
by placing the center of mass of the n books right over the edge of the bottom book
as in Figure 14.6.
So we know where to place the n C 1st book to get maximum overhang. In fact,
the reasoning above actually shows that this way of stacking n C 1 books is the
unique way to build a stable stack where the top book extends as far as possible.
All we have to do is calculate what this extension is.
“mcs” — 2017/6/5 — 19:42 — page 578 — #586
578
Chapter 14
Sums and Asymptotics
center of mass
of top n books
center of mass
of all n+1 books
}
top n books
}
1
2( n+1)
table
Figure 14.6 Additional overhang with n C 1 books.
The simplest way to do that is to let the center of mass of the top n books be the
origin. That way the horizontal coordinate of the center of mass of the whole stack
of n C 1 books will equal the increase in the overhang. But now the center of mass
of the bottom book has horizontal coordinate 1=2, so the horizontal coordinate of
center of mass of the whole stack of n C 1 books is
0 n C .1=2/ 1
1
D
:
nC1
2.n C 1/
In other words,
BnC1 D Bn C
1
;
2.n C 1/
(14.18)
as shown in Figure 14.6.
Expanding equation (14.18), we have
1
1
C
2n 2.n C 1/
1
1
1
D B1 C
C C
C
22
2n 2.n C 1/
nC1
1X1
D
:
2
i
BnC1 D Bn
1
C
i D1
So our next task is to examine the behavior of Bn as n grows.
(14.19)
“mcs” — 2017/6/5 — 19:42 — page 579 — #587
14.4. Hanging Out Over the Edge
14.4.2
579
Harmonic Numbers
Definition 14.4.1. The nth harmonic number Hn is
Hn WWD
n
X
1
i D1
i
:
So (14.19) means that
Hn
:
2
The first few harmonic numbers are easy to compute. For example, H4 D 1 C
1
1
25
1
C
2
3 C 4 D 12 > 2. The fact that H4 is greater than 2 has special significance:
it implies that the total extension of a 4-book stack is greater than one full book!
This is the situation shown in Figure 14.7.
Bn D
1/2
1/4
1/6
Table
1/8
Figure 14.7 Stack of four books with maximum overhang.
There is good news and bad news about harmonic numbers. The bad news is
that there is no known closed-form expression for the harmonic numbers. The
good news is that we can use Theorem 14.3.2 to get close upper and lower bounds
on Hn . In particular, since
Z n
ˇn
1
ˇ
dx D ln.x/ ˇ D ln.n/;
1
1 x
Theorem 14.3.2 means that
1
Hn ln.n/ C 1:
(14.20)
n
In other words, the nth harmonic number is very close to ln.n/.
Because the harmonic numbers frequently arise in practice, mathematicians have
worked hard to get even better approximations for them. In fact, it is now known
that
1
1
.n/
C
(14.21)
Hn D ln.n/ C C
C
2n 12n2
120n4
ln.n/ C
“mcs” — 2017/6/5 — 19:42 — page 580 — #588
580
Chapter 14
Sums and Asymptotics
Here is a value 0:577215664 : : : called Euler’s constant, and .n/ is between 0
and 1 for all n. We will not prove this formula.
We are now finally done with our analysis of the book stacking problem. Plugging the value of Hn into (14.19), we find that the maximum overhang for n books
is very close to ln.n/=2. Since ln.n/ grows to infinity as n increases, this means
that if we are given enough books we can get a book to hang out arbitrarily far
over the edge of the table. Of course, the number of books we need will grow as
an exponential function of the overhang; it will take 227 books just to achieve an
overhang of 3, never mind an overhang of 100.
Extending Further Past the End of the Table
The overhang we analyzed above was the furthest out the top book could extend
past the table. This leaves open the question of if there is some better way to build
a stable stack where some book other than the top stuck out furthest. For example,
Figure 14.8 shows a stable stack of two books where the bottom book extends
further out than the top book. Moreover, the bottom book extends 3/4 of a book
length past the end of the table, which is the same as the maximum overhang for
the top book in a two book stack.
Since the two book arrangement in Figure 14.8(a) ties the maximum overhang
stack in Figure 14.8(b), we could take the unique stable stack of n books where the
top book extends furthest, and switch the top two books to look like Figure 14.8(a).
This would give a stable stack of n books where the second from the top book
extends the same maximum overhang distance. So for n > 1, there are at least
two ways of building a stable stack of n books which both extend the maximum
overhang distance—one way where the top book is furthest out, and another way
where the second from the top book is furthest out.
It is not too hard to prove that these are the only two ways to get a stable stack
of books that achieves maximum overhang, providing we stick to stacking only
one book on top of another. But there is more to the story. Building book piles
with more than one book resting on another—think of an inverted pyramid—it is
p
possible to get a stack of n books to extend proportional to 3 n—much more than
ln n—book lengths without falling over. See [16], Maximum Overhang.
14.4.3
Asymptotic Equality
For cases like equation 14.21 where we understand the growth of a function like Hn
up to some (unimportant) error terms, we use a special notation, , to denote the
leading term of the function. For example, we say that Hn ln.n/ to indicate that
the leading term of Hn is ln.n/. More precisely:
“mcs” — 2017/6/5 — 19:42 — page 581 — #589
14.4. Hanging Out Over the Edge
581
table
1=2
3=4
(a)
table
1=4
1=2
(b)
Figure 14.8 Figure (a) shows a stable stack of two books where the bottom book
extends the same amount past the end of the table as the maximum overhang twobook stack shown in Figure (b).
“mcs” — 2017/6/5 — 19:42 — page 582 — #590
582
Chapter 14
Sums and Asymptotics
Definition 14.4.2. For functions f; g W R ! R, we say f is asymptotically equal
to g, in symbols,
f .x/ g.x/
iff
lim f .x/=g.x/ D 1:
x!1
Although it is tempting to write Hn ln.n/ C to indicate the two leading
terms, this is not really right. According to Definition 14.4.2, Hn ln.n/ C c
where c is any constant. The correct way to indicate that is the second-largest
term is Hn ln.n/ .
The reason that the notation is useful is that often we do not care about lower
order terms. For example, if n D 100, then we can compute H.n/ to great precision
using only the two leading terms:
ˇ
ˇ
ˇ 1
ˇ
1
1
ˇ
ˇ< 1 :
C
jHn ln.n/
jˇ
4
200 120000 120 100 ˇ 200
We will spend a lot more time talking about asymptotic notation at the end of the
chapter. But for now, let’s get back to using sums.
14.5
Products
We’ve covered several techniques for finding closed forms for sums but no methods
for dealing with products. Fortunately, we do not need to develop an entirely new
set of tools when we encounter a product such as
nŠ WWD
n
Y
i:
(14.22)
iD1
That’s because we can convert any product into a sum by taking a logarithm. For
example, if
n
Y
P D
f .i /;
i D1
then
ln.P / D
n
X
ln.f .i //:
i D1
We can then apply our summing tools to find a closed form (or approximate closed
form) for ln.P / and then exponentiate at the end to undo the logarithm.
“mcs” — 2017/6/5 — 19:42 — page 583 — #591
14.5. Products
583
For example, let’s see how this works for the factorial function nŠ. We start by
taking the logarithm:
ln.nŠ/ D ln.1 2 3 .n
1/ n/
D ln.1/ C ln.2/ C ln.3/ C C ln.n
D
n
X
1/ C ln.n/
ln.i /:
i D1
Unfortunately, no closed form for this sum is known. However, we can apply
Theorem 14.3.2 to find good closed-form bounds on the sum. To do this, we first
compute
Z n
ˇn
ˇ
ln.x/ dx D x ln.x/ x ˇ
1
1
D n ln.n/
n C 1:
Plugging into Theorem 14.3.2, this means that
n ln.n/
nC1 n
X
ln.i / n ln.n/
n C 1 C ln.n/:
i D1
Exponentiating then gives
nn
en
1
nŠ nnC1
:
en 1
This means that nŠ is within a factor of n of nn =e n
14.5.1
(14.23)
1.
Stirling’s Formula
The most commonly used product in discrete mathematics is probably nŠ, and
mathematicians have workedto find tight closed-form bounds on its value. The
most useful bounds are given in Theorem 14.5.1.
Theorem 14.5.1 (Stirling’s Formula). For all n 1,
n n
p
nŠ D 2 n
e .n/
e
where
1
1
.n/ :
12n C 1
12n
“mcs” — 2017/6/5 — 19:42 — page 584 — #592
584
Chapter 14
Sums and Asymptotics
Theorem 14.5.1 can be proved by induction (with some pain), and there are lots
of proofs using elementary calculus, but we won’t go into them.
There are several important things to notice about Stirling’s Formula. First, .n/
is always positive. This means that
n n
p
nŠ > 2 n
(14.24)
e
for all n 2 NC .
Second, .n/ tends to zero as n gets large. This means that
n n
p
nŠ 2 n
e
(14.25)
which is impressive. After all, who would expect both and e to show up in a
closed-form expression that is asymptotically equal to nŠ?
Third, .n/ is small even for small values of n. This means that Stirling’s Formula provides good approximations for nŠ for most all values of n. For example, if
we use
n n
p
2 n
e
as the approximation for nŠ, as many people do, we are guaranteed to be within a
factor of
1
e .n/ e 12n
of the correct value. For n 10, this means we will be within 1% of the correct
value. For n 100, the error will be less than 0.1%.
If we need an even closer approximation for nŠ, then we could use either
n n
p
e 1=12n
2 n
e
or
p
n n
e 1=.12nC1/
e
depending on whether we want an upper, or a lower, bound. By Theorem 14.5.1,
we know that both bounds will be within a factor of
2 n
1
e 12n
1
12nC1
1
D e 144n2 C12n
of the correct value. For n 10, this means that either bound will be within 0.01%
of the correct value. For n 100, the error will be less than 0.0001%.
For quick future reference, these facts are summarized in Corollary 14.5.2 and
Table 14.1.
“mcs” — 2017/6/5 — 19:42 — page 585 — #593
14.6. Double Trouble
585
Approximation
p
n
2 n ne
p
n
2 n ne e 1=12n
n1
n 10
n 100
n 1000
< 10%
< 1%
< 0.1%
< 0.01%
< 1%
< 0.01%
< 0.0001%
< 0.000001%
Table 14.1 Error bounds on common approximations
p
for nŠ from Theon n
rem 14.5.1. For example, if n 100, then 2 n e
approximates nŠ to
within 0.1%.
Corollary 14.5.2.
8
ˆ1:09
p
n n <
1:009
nŠ < 2 n
ˆ
e
:
1:0009
14.6
for n 1;
for n 10;
for n 100:
Double Trouble
Sometimes we have to evaluate sums of sums, otherwise known as double summations. This sounds hairy, and sometimes it is. But usually, it is straightforward—
you just evaluate the inner sum, replace it with a closed form, and then evaluate the
“mcs” — 2017/6/5 — 19:42 — page 586 — #594
586
Chapter 14
Sums and Asymptotics
outer sum (which no longer has a summation inside it). For example,5
!
1
n
1 X
X
X
x nC1
n
i
n1
y
equation 14.2
x D
y
1 x
nD0
nD0
i D0
D
D
D
X
1
1
1
x
y
n
1
nD0
.1
1
x/.1
.1
1
x/.1
x
x 1 x
y/
y/
.1
.1 xy/ x.1
.1 x/.1 y/.1
y/
xy/
D
.1
1 x
x/.1 y/.1
xy/
.1
1
y/.1
xy/
y n x nC1
nD0
1
X
.xy/n
Theorem 14.1.1
nD0
x
x/.1
D
D
X
1
1
xy/
Theorem 14.1.1
:
When there’s no obvious closed form for the inner sum, a special trick that is
often useful is to try exchanging the order of summation. For example, suppose we
want to compute the sum of the first n harmonic numbers
n
X
kD1
Hk D
n X
k
X
1
j
(14.26)
kD1 j D1
For intuition about this sum, we can apply Theorem 14.3.2 to equation 14.20 to
conclude that the sum is close to
Z n
ˇn
ˇ
ln.x/ dx D x ln.x/ x ˇ D n ln.n/ n C 1:
1
1
Now let’s look for an exact answer. If we think about the pairs .k; j / over which
5 OK, so maybe this one is a little hairy, but it is also fairly straightforward. Wait till you see the
next one!
“mcs” — 2017/6/5 — 19:42 — page 587 — #595
14.6. Double Trouble
587
we are summing, they form a triangle:
k 1
2
3
4
n
j
1
1
1
1
1
:::
1
2
3
4
5 :::
n
1=2
1=2 1=3
1=2 1=3 1=4
1=2
:::
1=n
The summation in equation 14.26 is summing each row and then adding the row
sums. Instead, we can sum the columns and then add the column sums. Inspecting
the table we see that this double sum can be written as
n
X
Hk D
n X
k
X
1
j
kD1 j D1
kD1
n
n X
X
1
D
j
j D1 kDj
D
n
n
X
1 X
1
j
j D1
D
n
X
j D1
D
kDj
1
.n
j
j C 1/
n
X
nC1
j
j D1
D .n C 1/
n
X
j
j
j D1
n
X
1
j
j D1
D .n C 1/Hn
n:
n
X
1
j D1
(14.27)
“mcs” — 2017/6/5 — 19:42 — page 588 — #596
588
14.7
Chapter 14
Sums and Asymptotics
Asymptotic Notation
Asymptotic notation is a shorthand used to give a quick measure of the behavior
of a function f .n/ as n grows large. For example, the asymptotic notation of
Definition 14.4.2 is a binary relation indicating that two functions grow at the same
rate. There is also a binary relation “little oh” indicating that one function grows at
a significantly slower rate than another and “Big Oh” indicating that one function
grows not much more rapidly than another.
14.7.1
Little O
Definition 14.7.1. For functions f; g W R ! R, with g nonnegative, we say f is
asymptotically smaller than g, in symbols,
f .x/ D o.g.x//;
iff
lim f .x/=g.x/ D 0:
x!1
For example, 1000x 1:9 D o.x 2 / because 1000x 1:9 =x 2 D 1000=x 0:1 and since
goes to infinity with x and 1000 is constant, we have limx!1 1000x 1:9 =x 2 D
0. This argument generalizes directly to yield
x 0:1
Lemma 14.7.2. x a D o.x b / for all nonnegative constants a < b.
Using the familiar fact that log x < x for all x > 1, we can prove
Lemma 14.7.3. log x D o.x / for all > 0.
Proof. Choose > ı > 0 and let x D z ı in the inequality log x < x. This implies
log z < z ı =ı D o.z /
by Lemma 14.7.2:
(14.28)
Corollary 14.7.4. x b D o.ax / for any a; b 2 R with a > 1.
Lemma 14.7.3 and Corollary 14.7.4 can also be proved using l’Hôpital’s Rule or
the Maclaurin Series for log x and e x . Proofs can be found in most calculus texts.
“mcs” — 2017/6/5 — 19:42 — page 589 — #597
14.7. Asymptotic Notation
14.7.2
589
Big O
“Big Oh” is the most frequently used asymptotic notation. It is used to give an
upper bound on the growth of a function, such as the running time of an algorithm.
There is a standard definition of Big Oh given below in 14.7.9, but we’ll begin with
an alternative definition that makes apparent several basic properties of Big Oh.
Definition 14.7.5. Given functions f; g W R ! R with g nonnegative, we say that
f D O.g/
iff
lim sup jf .x/j =g.x/ < 1:
x!1
Here we’re using the technical notion of limit superior6 instead of just limit. But
because limits and lim sup’s are the same when limits exist, this formulation makes
it easy to check basic properties of Big Oh. We’ll take the following Lemma for
granted.
Lemma 14.7.6. If a function f W R ! R has a finite or infinite limit as its argument
approaches infinity, then its limit and limit superior are the same.
Now Definition 14.7.5 immediately implies:
Lemma 14.7.7. If f D o.g/ or f g, then f D O.g/.
Proof. lim f =g D 0 or lim f =g D 1 implies lim f =g < 1, so by Lemma 14.7.6,
lim sup f =g < 1.
Note that the converse of Lemma 14.7.7 is not true. For example, 2x D O.x/,
but 2x 6 x and 2x ¤ o.x/.
We also have:
Lemma 14.7.8. If f D o.g/, then it is not true that g D O.f /.
Proof.
g.x/
1
1
D
D D 1;
x!1 f .x/
limx!1 f .x/=g.x/
0
lim
so by Lemma 14.7.6, g ¤ O.f /.
6 The
precise definition of lim sup is
lim sup h.x/ WWD lim lubyx h.y/;
x!1
where “lub” abbreviates “least upper bound.”
x!1
“mcs” — 2017/6/5 — 19:42 — page 590 — #598
590
Chapter 14
Sums and Asymptotics
We need lim sup’s in Definition 14.7.5 to cover cases when limits don’t exist. For
example, if f .x/=g.x/ oscillates between 3 and 5 as x grows, then limx!1 f .x/=g.x/
does not exist, but f D O.g/ because lim supx!1 f .x/=g.x/ D 5.
An equivalent, more usual formulation of big O does not mention lim sup’s:
Definition 14.7.9. Given functions f; g W R ! R with g nonnegative, we say
f D O.g/
iff there exists a constant c 0 and an x0 such that for all x x0 , jf .x/j cg.x/.
This definition is rather complicated, but the idea is simple: f .x/ D O.g.x//
means f .x/ is less than or equal to g.x/, except that we’re willing to ignore a
constant factor, namely c and to allow exceptions for small x, namely, x < x0 . So
in the case that f .x/=g.x/ oscillates between 3 and 5, f D O.g/ according to
Definition 14.7.9 because f 5g.
Proposition 14.7.10. 100x 2 D O.x 2 /.
Proof. ˇChooseˇ c D 100 and x0 D 1. Then the proposition holds, since for all
x 1, ˇ100x 2 ˇ 100x 2 .
Proposition 14.7.11. x 2 C 100x C 10 D O.x 2 /.
Proof. .x 2 C100x C10/=x 2 D 1C100=x C10=x 2 and so its limit as x approaches
infinity is 1C0C0 D 1. So in fact, x 2 C100xC10 x 2 , and therefore x 2 C100xC
10 D O.x 2 /. Indeed, it’s conversely true that x 2 D O.x 2 C 100x C 10/.
Proposition 14.7.11 generalizes to an arbitrary polynomial:
Proposition 14.7.12. ak x k C ak
1x
k 1
C C a1 x C a0 D O.x k /.
We’ll omit the routine proof.
Big O notation is especially useful when describing the running time of an algorithm. For example, the usual algorithm for multiplying n n matrices uses a
number of operations proportional to n3 in the worst case. This fact can be expressed concisely by saying that the running time is O.n3 /. So this asymptotic
notation allows the speed of the algorithm to be discussed without reference to
constant factors or lower-order terms that might be machine specific. It turns out
that there is another matrix multiplication procedure that uses O.n2:55 / operations.
The fact that this procedure is asymptotically faster indicates that it involves new
ideas that go beyond a simply more efficient implementation of the O.n3 / method.
Of course the asymptotically faster procedure will also definitely be much more
efficient on large enough matrices, but being asymptotically faster does not mean
“mcs” — 2017/6/5 — 19:42 — page 591 — #599
14.7. Asymptotic Notation
591
that it is a better choice. The O.n2:55 /-operation multiplication procedure is almost
never used in practice because it only becomes more efficient than the usual O.n3 /
procedure on matrices of impractical size.7
14.7.3
Theta
Sometimes we want to specify that a running time T .n/ is precisely quadratic up to
constant factors (both upper bound and lower bound). We could do this by saying
that T .n/ D O.n2 / and n2 D O.T .n//, but rather than say both, mathematicians
have devised yet another symbol ‚ to do the job.
Definition 14.7.13.
f D ‚.g/ iff
f D O.g/ and g D O.f /:
The statement f D ‚.g/ can be paraphrased intuitively as “f and g are equal
to within a constant factor.”
The Theta notation allows us to highlight growth rates and suppress distracting
factors and low-order terms. For example, if the running time of an algorithm is
T .n/ D 10n3
20n2 C 1;
then we can more simply write
T .n/ D ‚.n3 /:
In this case, we would say that T is of order n3 or that T .n/ grows cubically, which
is often the main thing we really want to know. Another such example is
2 3x
7
C
.2:7x 113 C x 9
p
x
86/4
1:083x D ‚.3x /:
Just knowing that the running time of an algorithm is ‚.n3 /, for example, is
useful, because if n doubles we can predict that the running time will by and large8
increase by a factor of at most 8 for large n. In this way, Theta notation preserves information about the scalability of an algorithm or system. Scalability is, of course,
a big issue in the design of algorithms and systems.
7 It is even conceivable that there is an O.n2 / matrix multiplication procedure, but none is known.
8 Since ‚.n3 / only implies that the running time T .n/ is between cn3 and d n3 for constants
0 < c < d , the time T .2n/ could regularly exceed T .n/ by a factor as large as 8d=c. The factor is
sure to be close to 8 for all large n only if T .n/ n3 .
“mcs” — 2017/6/5 — 19:42 — page 592 — #600
592
Chapter 14
14.7.4
Sums and Asymptotics
Pitfalls with Asymptotic Notation
There is a long list of ways to make mistakes with asymptotic notation. This section
presents some of the ways that big O notation can lead to trouble. With minimal
effort, you can cause just as much chaos with the other symbols.
The Exponential Fiasco
Sometimes relationships involving big O are not so obvious. For example, one
might guess that 4x D O.2x / since 4 is only a constant factor larger than 2. This
reasoning is incorrect, however; 4x actually grows as the square of 2x .
Constant Confusion
Every constant is O.1/. For example, 17 D O.1/. This is true because if we let
f .x/ D 17 and g.x/ D 1, then there exists a c > 0 and an x0 such that jf .x/j cg.x/. In particular, we could choose c = 17 and x0 D 1, since j17j 17 1 for all
x 1. We can construct a false theorem that exploits this fact.
False Theorem 14.7.14.
n
X
i D O.n/
i D1
P
Bogus proof. Define f .n/ D niD1 i D 1 C 2 C 3 C C n. Since we have shown
that every constant i is O.1/, f .n/ D O.1/ C O.1/ C C O.1/ D O.n/.
Pn
Of course in reality i D1 i D n.n C 1/=2 ¤ O.n/.
The error stems from confusion over what is meant in the statement i D O.1/.
For any constant i 2 N it is true that i D O.1/. More precisely, if f is any constant
function, then f D O.1/. But in this False Theorem, i is not constant—it ranges
over a set of values 0; 1; : : : ; n that depends on n.
And anyway, we should not be adding O.1/’s as though they were numbers. We
never even defined what O.g/ means by itself; it should only be used in the context
“f D O.g/” to describe a relation between functions f and g.
Equality Blunder
The notation f D O.g/ is too firmly entrenched to avoid, but the use of “=” is
regrettable. For example, if f D O.g/, it seems quite reasonable to write O.g/ D
f . But doing so might tempt us to the following blunder: because 2n D O.n/, we
can say O.n/ D 2n. But n D O.n/, so we conclude that n D O.n/ D 2n, and
therefore n D 2n. To avoid such nonsense, we will never write “O.f / D g.”
“mcs” — 2017/6/5 — 19:42 — page 593 — #601
14.7. Asymptotic Notation
593
Similarly, you will often see statements like
Hn D ln.n/ C
1
CO
n
or
n n
p
nŠ D .1 C o.1// 2 n
e
In such cases, the true meaning is
Hn D ln.n/ C
C f .n/
for some f .n/ where f .n/ D O.1=n/, and
n n
p
nŠ D .1 C g.n// 2 n
e
where g.n/ D o.1/. These last transgressions are OK as long as you (and your
reader) know what you mean.
Operator Application Blunder
It’s tempting to assume that familiar operations preserve asymptotic relations, but
it ain’t necessarily so. For example, f g in general does not even imply that
3f D ‚ .3g /. On the other hand, some operations preserve and even strengthen
asymptotic relations, for example,
f D ‚.g/ IMPLIES ln f ln g:
See Problem 14.27.
14.7.5
Omega (Optional)
Sometimes people incorrectly use Big Oh in the context of a lower bound. For
example, they might say, “The running time T .n/ is at least O.n2 /.” This is another
blunder! Big Oh can only be used for upper bounds. The proper way to express the
lower bound would be
n2 D O.T .n//:
The lower bound can also be described with another special notation “big Omega.”
Definition 14.7.15. Given functions f; g W R ! R with f nonnegative, define
f D .g/
to mean
g D O.f /:
“mcs” — 2017/6/5 — 19:42 — page 594 — #602
594
Chapter 14
Sums and Asymptotics
p
For example, x 2 D .x/, 2x D .x 2 / and x=100 D .100x C x/.
So if the running time of your algorithm on inputs of size n is T .n/, and you
want to say it is at least quadratic, say
T .n/ D .n2 /:
There is a similar “little omega” notation for lower bounds corresponding to little
o:
Definition 14.7.16. For functions f; g W R ! R with f nonnegative, define
f D !.g/
to mean
g D o.f /:
p
For example, x 1:5 D !.x/ and x D !.ln2 .x//.
The little omega symbol is not as widely used as the other asymptotic symbols
we defined.
Problems for Section 14.1
Class Problems
Problem 14.1.
We begin with two large glasses. The first glass contains a pint of water, and the
second contains a pint of wine. We pour 1/3 of a pint from the first glass into the
second, stir up the wine/water mixture in the second glass, and then pour 1/3 of
a pint of the mix back into the first glass and repeat this pouring back-and-forth
process a total of n times.
(a) Describe a closed-form formula for the amount of wine in the first glass after
n back-and-forth pourings.
(b) What is the limit of the amount of wine in each glass as n approaches infinity?
Problem 14.2.
“mcs” — 2017/6/5 — 19:42 — page 595 — #603
14.7. Asymptotic Notation
595
You’ve seen this neat trick for evaluating a geometric sum:
S D 1 C z C z2 C : : : C zn
zS D z C z 2 C : : : C z n C z nC1
zS D 1
S
SD
1
z nC1
z nC1
(where z ¤ 1/
1 z
Use the same approach to find a closed-form expression for this sum:
T D 1z C 2z 2 C 3z 3 C : : : C nz n
Problem 14.3.
Sammy the Shark is a financial service provider who offers loans on the following
terms.
Sammy loans a client m dollars in the morning. This puts the client m dollars
in debt to Sammy.
Each evening, Sammy first charges a service fee which increases the client’s
debt by f dollars, and then Sammy charges interest, which multiplies the
debt by a factor of p. For example, Sammy might charge a “modest” ten
cent service fee and 1% interest rate per day, and then f would be 0:1 and p
would be 1:01.
(a) What is the client’s debt at the end of the first day?
(b) What is the client’s debt at the end of the second day?
(c) Write a formula for the client’s debt after d days and find an equivalent closed
form.
(d) If you borrowed $10 from Sammy for a year, how much would you owe him?
Homework Problems
Problem 14.4.
Is a Harvard degree really worth more than an MIT degree? Let us say that a person
with a Harvard degree starts with $40,000 and gets a $20,000 raise every year after
graduation, whereas a person with an MIT degree starts with $30,000, but gets a
20% raise every year. Assume inflation is a fixed 8% every year. That is, $1.08 a
year from now is worth $1.00 today.
“mcs” — 2017/6/5 — 19:42 — page 596 — #604
596
Chapter 14
Sums and Asymptotics
(a) How much is a Harvard degree worth today if the holder will work for n years
following graduation?
(b) How much is an MIT degree worth in this case?
(c) If you plan to retire after twenty years, which degree would be worth more?
Problem 14.5.
Suppose you deposit $100 into your MIT Credit Union account today, then$99 at
the end of the first month from now, $98 at tne end of the second months from now,
and so on. Given that the interest rate is constantly 0.3% per month, how long will
it take to save $5,000?
Problems for Section 14.2
Class Problems
Problem 14.6.
Find a closed form for each of the following sums:
(a)
n X
1
1
:
i C 2012 i C 2013
i D1
(b) Assuming the following sum equals a polynomial in n, find the polynomial.
Then verify by induction that the sum equals the polynomial you find.
n
X
i3
i D1
Problems for Section 14.3
Practice Problems
Problem 14.7.
Let
5
X
p
S WWD
3n :
nD1
“mcs” — 2017/6/5 — 19:42 — page 597 — #605
14.7. Asymptotic Notation
597
Using the Integral Method of Section 14.3, we can find integers a, b, c, d and a
real number e such that
Z b
Z d
e
x dx S x e dx
a
c
What are appropriate values for a; : : : ; e?
Class Problems
Problem 14.8.
Let f W R ! R be a continuous, weakly increasing function. Say that f grows
slowly when
Z n
f .n/ D o
f .x/ dx :
(a) Prove that the function fa .n/ WWD
1
a
n grows
slowly for any a > 0.
(b) Prove that the function e n does not grow slowly.
(c) Prove that if f grows slowly, then
n
Z
f .x/ dx 1
n
X
f .i / :
i D1
Exam Problems
Problem 14.9.
Assume n is an integer larger than 1. Circle all the correct inequalities below.
Explanations are not required, but partial credit for wrong answers will not be
given without them. Hint: You may find the graphs in Figure 14.9 helpful.
n
X
Z
ln.i C 1/ ln 2 C
i D1
n
X
Z
ln.i C 1/ i D1
n
X
1
i D1
i
Z
0
n
0
1
n
ln.x C 1/dx
n
ln.x C 2/dx
1
dx
xC1
“mcs” — 2017/6/5 — 19:42 — page 598 — #606
598
Chapter 14
Sums and Asymptotics
2.5
y = ln(x+1)
2
1.5
y = ln(x+2)
1
0.5
0
0
1
2
3
4
5
6
7
8
7
8
1
0.8
y = 1/x
0.6
y = 1/(x+1)
0.4
0.2
0
0
1
2
3
4
5
6
Figure 14.9 Integral bounds for two sums
“mcs” — 2017/6/5 — 19:42 — page 599 — #607
14.7. Asymptotic Notation
599
Homework Problems
Problem 14.10.
Let f W RC ! RC be a weakly decreasing function. Define
S WWD
n
X
f .i /
i D1
and
Z
I WWD
n
f .x/ dx:
1
Prove that
I C f .n/ S I C f .1/:
(Proof by very clear picture is OK.)
Problem 14.11.
Use integration to find upper and lower bounds that differ by at most 0.1 for the
following sum. (You may need to add the first few terms explicitly and then use
integrals to bound the sum of the remaining terms.)
1
X
iD1
1
.2i C 1/2
Problems for Section 14.4
Class Problems
Problem 14.12.
An explorer is trying to reach the Holy Grail, which she believes is located in a
desert shrine d days walk from the nearest oasis. In the desert heat, the explorer
must drink continuously. She can carry at most 1 gallon of water, which is enough
for 1 day. However, she is free to make multiple trips carrying up to a gallon each
time to create water caches out in the desert.
For example, if the shrine were 2=3 of a day’s walk into the desert, then she could
recover the Holy Grail after two days using the following strategy. She leaves the
oasis with 1 gallon of water, travels 1=3 day into the desert, caches 1=3 gallon, and
then walks back to the oasis—arriving just as her water supply runs out. Then she
picks up another gallon of water at the oasis, walks 1=3 day into the desert, tops off
“mcs” — 2017/6/5 — 19:42 — page 600 — #608
600
Chapter 14
Sums and Asymptotics
her water supply by taking the 1=3 gallon in her cache, walks the remaining 1=3
day to the shrine, grabs the Holy Grail, and then walks for 2=3 of a day back to the
oasis—again arriving with no water to spare.
But what if the shrine were located farther away?
(a) What is the most distant point that the explorer can reach and then return to
the oasis, with no water precached in the desert, if she takes a total of only 1 gallon
from the oasis?
(b) What is the most distant point the explorer can reach and still return to the
oasis if she takes a total of only 2 gallons from the oasis? No proof is required; just
do the best you can.
(c) The explorer will travel using a recursive strategy to go far into the desert and
back, drawing a total of n gallons of water from the oasis. Her strategy is to build
up a cache of n 1 gallons, plus enough to get home, a certain fraction of a day’s
distance into the desert. On the last delivery to the cache, instead of returning home,
she proceeds recursively with her n 1 gallon strategy to go farther into the desert
and return to the cache. At this point, the cache has just enough water left to get
her home.
Prove that with n gallons of water, this strategy will get her Hn =2 days into the
desert and back, where Hn is the nth Harmonic number:
Hn WWD
1 1 1
1
C C C C :
1 2 3
n
Conclude that she can reach the shrine, however far it is from the oasis.
(d) Suppose that the shrine is d D 10 days walk into the desert. Use the asymptotic approximation Hn ln n to show that it will take more than a million years
for the explorer to recover the Holy Grail.
Problem 14.13.
P
p
There is a number a such that 1
i D1 i converges iff p < a. What is the value of
a? Prove it.
Hint: Find a value for a you think that works, then apply the integral bound.
Problem 14.14.
An infinite sum of nonnegative terms will converge to the same value—or diverge—
no matter the order in which the terms are summed. This may not be true when
“mcs” — 2017/6/5 — 19:42 — page 601 — #609
14.7. Asymptotic Notation
601
there are an infinite number of both nonnegative and negative terms. An extreme
example is
1
X
. 1/i D 1 C . 1/ C 1 C . 1/ C i D0
because by regrouping the terms we can deduce:
Œ1 C . 1/ C Œ1 C . 1/ C D 0 C 0 C D 0;
1 C Œ. 1/ C 1 C Œ. 1/ C 1 C D 1 C 0 C 0 C D 1:
The problem here with this infinite sum is that the sum of the first n terms oscillates between 0 and 1, so the sum does not approach any limit.
But even for convergent sums, rearranging terms can cause big changes when
the sum contains positive and negative terms. To illustrate the problem, we look at
the Alternating Harmonic Series:
1
1=2 C 1=3
1=4 C ˙ :
A standard result of elementary calculus, [2], p.403, is that this series converges to
ln 2, but things change if we reorder the terms in the series.
Explain for example how to reorder terms in the Alternating Harmonic Series so
that the reordered series converges to 7. Then explain how to reorder so it diverges.
Homework Problems
Problem 14.15.
There is a bug on the edge of a 1-meter rug. The bug wants to cross to the other
side of the rug. It crawls at 1 cm per second. However, at the end of each second,
a malicious first-grader named Mildred Anderson stretches the rug by 1 meter. Assume that her action is instantaneous and the rug stretches uniformly. Thus, here’s
what happens in the first few seconds:
The bug walks 1 cm in the first second, so 99 cm remain ahead.
Mildred stretches the rug by 1 meter, which doubles its length. So now there
are 2 cm behind the bug and 198 cm ahead.
The bug walks another 1 cm in the next second, leaving 3 cm behind and 197
cm ahead.
Then Mildred strikes, stretching the rug from 2 meters to 3 meters. So there
are now 3 .3=2/ D 4:5 cm behind the bug and 197 .3=2/ D 295:5 cm
ahead.
“mcs” — 2017/6/5 — 19:42 — page 602 — #610
602
Chapter 14
Sums and Asymptotics
The bug walks another 1 cm in the third second, and so on.
Your job is to determine this poor bug’s fate.
(a) During second i , what fraction of the rug does the bug cross?
(b) Over the first n seconds, what fraction of the rug does the bug cross altogether?
Express your answer in terms of the Harmonic number Hn .
(c) The known universe is thought to be about 3 1010 light years in diameter.
How many universe diameters must the bug travel to get to the end of the rug?
(This distance is NOT the inflated distance caused by the stretching but only the
actual walking done by the bug).
Problem 14.16.
Prove that the Alternating Harmonic Series
1
1=2 C 1=3
1=4 C ˙
converges.
Exam Problems
Problem 14.17.
Show that
1
X
ip
i D1
converges to a finite value iff p <
1.
Problems for Section 14.7
Practice Problems
Problem 14.18.
Find the least nonnegative integer n such that f .x/ is O.x n / when f is defined by
each of the expressions below.
(a) 2x 3 C .log x/x 2
(b) 2x 2 + .log x/x 3
“mcs” — 2017/6/5 — 19:42 — page 603 — #611
14.7. Asymptotic Notation
603
(c) .1:1/x
(d) .0:1/x
(e) .x 4 C x 2 C 1/=.x 3 C 1/
(f) .x 4 C 5 log x/=.x 4 C 1/
(g) 2.3 log2 x
2/
Problem 14.19.
Let f .n/ D n3 . For each function g.n/ in the table below, indicate which of the
indicated asymptotic relations hold.
g.n/
f D O.g/ f D o.g/ g D O.f / g D o.f /
2
3
6 5n 4n C 3n
n3 log n
.sin . n=2/ C 2/ n3
nsin. n=2/C2
log nŠ
e 0:2n 100n3
Problem 14.20.
Circle each of the true statements below.
Explanations are not required, but partial credit for wrong answers will not be
given without them.
n2 n2 C n
3n D O 2n
nsin.n=2/C1 D o n2
3n3
nD‚
.n C 1/.n 1/
Problem 14.21.
Show that
ln.n2 Š/ D ‚.n2 ln n/
“mcs” — 2017/6/5 — 19:42 — page 604 — #612
604
Chapter 14
Sums and Asymptotics
Hint: Stirling’s formula for .n2 /Š.
Problem 14.22.
The quantity
.2n/Š
22n .nŠ/2
(14.29)
will come up later in the course (it is the probability that in 22n flips of a fair coin,
1
exactly n will be Heads). Show that it is asymptotically equal to p .
n
Problem 14.23.
Suppose let f and g be real-valued functions.
(a) Give an example of f; g such that
lim sup fg < lim sup f lim sup g;
and all the lim sup’s are finite.
(b) Give an example of f; g such that
lim sup fg > lim sup f lim sup g:
and all the lim sup’s are finite.
Homework Problems
Problem 14.24. (a) Prove that log x < x for all x > 1 (requires elementary calculus).
(b) Prove that the relation R on functions such that f R g iff g D o.f / is a strict
partial order.
(c) Prove that f g iff f D g C h for some function h D o.g/.
Problem 14.25.
Indicate which of the following holds for each pair of functions .f .n/; g.n// in
the table below. Assume k 1, > 0, and c > 1 are constants. Pick the four
“mcs” — 2017/6/5 — 19:42 — page 605 — #613
14.7. Asymptotic Notation
605
table entries you consider to be the most challenging or interesting and justify your
answers to these.
f .n/
g.n/
2n
2n=2
f D O.g/ f D o.g/ g D O.f / g D o.f / f D ‚.g/ f g
p
n
nsin.n=2/
log.nŠ/ log.nn /
nk
cn
logk n
n
Problem 14.26.
Arrange the following functions in a sequence f1 , f2 , ... f24 so that fi D O.fi C1 /.
Additionally, if fi D ‚.fi C1 /, indicate that too:
1. n log n
2. 2100 n
3. n
1
4. n
1=2
5. .log n/=n
n
6. 64
7. nŠ
8. 22
100
9. 22
n
10. 2n
11. 3n
12. n2n
13. 2nC1
14. 2n
15. 3n
“mcs” — 2017/6/5 — 19:42 — page 606 — #614
606
Chapter 14
Sums and Asymptotics
16. log .nŠ/
17. log2 n
18. log10 n
19. 2:1
p
n
20. 22n
21. 4n
22. n64
23. n65
24. nn
Problem 14.27.
Let f , g be nonnegative real-valued functions such that limx!1 f .x/ D 1 and
f g.
(a) Give an example of f; g such that NOT.2f 2g /.
(b) Prove that log f log g.
(c) Use Stirling’s formula to prove that in fact
log.nŠ/ n log n
Problem 14.28.
Determine which of these choices
‚.n/;
‚.n2 log n/;
‚.n2 /;
‚.1/;
‚.2n /;
‚.2n ln n /;
none of these
describes each function’s asymptotic behavior. Full proofs are not required, but
briefly explain your answers.
(a)
n C ln n C .ln n/2
(b)
n2 C 2n 3
n2 7
“mcs” — 2017/6/5 — 19:42 — page 607 — #615
14.7. Asymptotic Notation
607
(c)
n
X
22iC1
i D0
(d)
ln.n2 Š/
(e)
n
X
k 1
kD1
1
2k
Problem 14.29. (a) Either prove or disprove each of the following statements.
nŠ D O..n C 1/Š/
.n C 1/Š D O.nŠ/
nŠ D ‚..n C 1/Š/
nŠ D o..n C 1/Š/
.n C 1/Š D o.nŠ/
nCe
D o.nŠ/.
(b) Show that n3
Problem 14.30.
Prove that
n
X
k 6 D ‚.n7 /:
kD1
Class Problems
Problem 14.31.
Give an elementary proof (without appealing to Stirling’s formula) that log.nŠ/ D
‚.n log n/.
Problem 14.32.
Suppose f; g W NC ! NC and f g.
(a) Prove that 2f 2g.
“mcs” — 2017/6/5 — 19:42 — page 608 — #616
608
Chapter 14
Sums and Asymptotics
(b) Prove that f 2 g 2 .
(c) Give examples of f and g such that 2f 6 2g .
Problem 14.33.
Recall that for functions f; g on N, f D O.g/ iff
9c 2 N 9n0 2 N 8n n0
c g.n/ jf .n/j :
(14.30)
For each pair of functions below, determine whether f D O.g/ and whether
g D O.f /. In cases where one function is O() of the other, indicate the smallest
nonnegative integer c and for that smallest c, the smallest corresponding nonnegative integer n0 ensuring that condition (14.30) applies.
(a) f .n/ D n2 ; g.n/ D 3n.
f D O.g/
YES
NO
If YES, c D
, n0 =
g D O.f /
YES
NO
If YES, c D
, n0 =
(b) f .n/ D .3n
7/=.n C 4/; g.n/ D 4
f D O.g/
YES
NO
If YES, c D
, n0 =
g D O.f /
YES
NO
If YES, c D
, n0 =
(c) f .n/ D 1 C .n sin.n=2//2 ; g.n/ D 3n
f D O.g/
YES
NO
If yes, c D
n0 =
g D O.f /
YES
NO
If yes, c D
n0 =
Problem 14.34.
False Claim.
2n D O.1/:
(14.31)
Explain why the claim is false. Then identify and explain the mistake in the
following bogus proof.
Bogus proof. The proof is by induction on n where the induction hypothesis P .n/
is the assertion (14.31).
base case: P .0/ holds trivially.
“mcs” — 2017/6/5 — 19:42 — page 609 — #617
14.7. Asymptotic Notation
2n
609
inductive step: We may assume P .n/, so there is a constant c > 0 such that
c 1. Therefore,
2nC1 D 2 2n .2c/ 1;
which implies that 2nC1 D O.1/. That is, P .n C 1/ holds, which completes the
proof of the inductive step.
We conclude by induction that 2n D O.1/ for all n. That is, the exponential
function is bounded by a constant.
Problem 14.35. (a) Prove that the relation R on functions such that f R g iff
f D o.g/ is a strict partial order.
(b) Describe two functions f; g that are incomparable under big Oh:
f ¤ O.g/ AND g ¤ O.f /:
Conclude that R is not a linear order. How about three such functions?
Exam Problems
Problem 14.36.
Give an example of a pair of strictly increasing total functions, f W NC ! NC and
g W NC ! NC , that satisfy f g but not 3f D O .3g /.
Problem 14.37.
Let f; g be real-valued functions such that f D ‚.g/ and limx!1 f .x/ D 1.
Prove that
ln f ln g:
Problem 14.38. (a) Show that
.an/b=n 1:
where a; b are positive constants and denotes asymptotic equality. Hint: an D
a2log2 n .
(b) You may assume that if f .n/ 1 and g.n/ 1 for all n, then f g !
1
1
f n g n . Show that
p
n
nŠ D ‚.n/:
“mcs” — 2017/6/5 — 19:42 — page 610 — #618
610
Chapter 14
Sums and Asymptotics
Problem 14.39.
(a) Define a function f .n/ such that f D ‚.n2 / and NOT.f n2 /.
f .n/ D
(b) Define a function g.n/ such that g D O.n2 /, g ¤ ‚.n2 /, g ¤ o.n2 /, and
n D O.g/.
g.n/ D
Problem 14.40. (a) Show that
.an/b=n 1:
where a; b are positive constants and denotes asymptotic equality. Hint: an D
a2log2 n .
(b) Show that
p
n
nŠ D ‚.n/:
Problem 14.41.
(a) Indicate which of the following asymptotic relations below on the set of nonnegative real-valued functions are equivalence relations (E), strict partial orders (S),
weak partial orders (W), or none of the above (N).
f g, the “asymptotically equal” relation.
f D o.g/, the “little Oh” relation.
f D O.g/, the “big Oh” relation.
f D ‚.g/, the “Theta” relation.
f D O.g/ AND NOT.g D O.f //.
(b) Indicate the implications among the assertions in part (a). For example,
f D o.g/ IMPLIES f D O.g/:
“mcs” — 2017/6/5 — 19:42 — page 611 — #619
14.7. Asymptotic Notation
611
Problem 14.42.
Recall that if f and g are nonnegative real-valued functions on ZC , then f D O.g/
iff there exist c; n0 2 ZC such that
8n n0 : f .n/ cg.n/:
For each pair of functions f and g below, indicate the smallest c 2 ZC , and
for that smallest c, the smallest corresponding n0 2 ZC , that would establish
f D O.g/ by the definition given above. If there is no such c, write 1.
(a) f .n/ D
1
2
ln n2 ; g.n/ D n.
cD
, n0 =
(b) f .n/ D n; g.n/ D n ln n.
cD
, n0 =
(c) f .n/ D 2n ; g.n/ D n4 ln n
.n 1/
C 2; g.n/ D 0:2.
(d) f .n/ D 3 sin
100
cD
, n0 =
cD
, n0 =
Problem 14.43.
Let f; g be positive real-valued functions on finite, connected, simple graphs. We
will extend the O./ notation to such graph functions as follows: f D O.g/ iff
there is a constant c > 0 such that
f .G/ c g.G/ for all connected simple graphs G with more than one vertex:
For each of the following assertions, state whether it is True or False and briefly
explain your answer. You are not expected to offer a careful proof or detailed
counterexample.
Reminder: V .G/ is the set of vertices and E.G/ is the set of edges of G, and G
is connected.
(a) jV .G/j D O.jE.G/j/.
(b) jE.G/j D O.jV .G/j/.
(c) jV .G/j D O..G//, where .G/ is the chromatic number of G.
(d) .G/ D O.jV .G/j/.
“mcs” — 2017/6/5 — 19:42 — page 612 — #620
“mcs” — 2017/6/5 — 19:42 — page 613 — #621
15
15.1
Cardinality Rules
Counting One Thing by Counting Another
How do you count the number of people in a crowded room? You could count
heads, since for each person there is exactly one head. Alternatively, you could
count ears and divide by two. Of course, you might have to adjust the calculation if
someone lost an ear in a pirate raid or someone was born with three ears. The point
here is that you can often count one thing by counting another, though some fudging may be required. This is a central theme of counting, from the easiest problems
to the hardest. In fact, we’ve already seen this technique used in Theorem 4.5.5,
where the number of subsets of an n-element set was proved to be the same as the
number of length-n bit-strings, by describing a bijection between the subsets and
the bit-strings.
The most direct way to count one thing by counting another is to find a bijection
between them, since if there is a bijection between two sets, then the sets have the
same size. This important fact is commonly known as the Bijection Rule. We’ve
already seen it as the Mapping Rules bijective case (4.7).
15.1.1
The Bijection Rule
The Bijection Rule acts as a magnifier of counting ability; if you figure out the size
of one set, then you can immediately determine the sizes of many other sets via
bijections. For example, let’s look at the two sets mentioned at the beginning of
Part III:
A D all ways to select a dozen donuts when five varieties are available
B D all 16-bit sequences with exactly 4 ones
An example of an element of set A is:
00
„ƒ‚…
chocolate
„ƒ‚…
lemon-filled
0„ 0 ƒ‚
0 0 0 …0
sugar
00
„ƒ‚…
glazed
00
„ƒ‚…
plain
Here, we’ve depicted each donut with a 0 and left a gap between the different
varieties. Thus, the selection above contains two chocolate donuts, no lemon-filled,
six sugar, two glazed, and two plain. Now let’s put a 1 into each of the four gaps:
00
„ƒ‚…
chocolate
1
1
„ƒ‚…
lemon-filled
0 0 ƒ‚
0 0 0 …0
„
sugar
1
00
„ƒ‚…
glazed
1
00
„ƒ‚…
plain
“mcs” — 2017/6/5 — 19:42 — page 614 — #622
614
Chapter 15
Cardinality Rules
and close up the gaps:
0011000000100100 :
We’ve just formed a 16-bit number with exactly 4 ones—an element of B!
This example suggests a bijection from set A to set B: map a dozen donuts
consisting of:
c chocolate, l lemon-filled, s sugar, g glazed, and p plain
to the sequence:
: : : 0…
„0 ƒ‚
1
c
: : : 0…
„0 ƒ‚
l
1
: : : 0…
„0 ƒ‚
s
1
: : : 0…
„0 ƒ‚
g
1
: : : 0…
„0 ƒ‚
p
The resulting sequence always has 16 bits and exactly 4 ones, and thus is an
element of B. Moreover, the mapping is a bijection: every such bit sequence comes
from exactly one order of a dozen donuts. Therefore, jAj D jBj by the Bijection
Rule. More generally,
Lemma 15.1.1. The number of ways to select n donuts when k flavors are available
is the same as the number of binary sequences with exactly n zeroes and k 1 ones.
This example demonstrates the power of the bijection rule. We managed to prove
that two very different sets are actually the same size—even though we don’t know
exactly how big either one is. But as soon as we figure out the size of one set, we’ll
immediately know the size of the other.
This particular bijection might seem frighteningly ingenious if you’ve not seen
it before. But you’ll use essentially this same argument over and over, and soon
you’ll consider it routine.
15.2
Counting Sequences
The Bijection Rule lets us count one thing by counting another. This suggests a
general strategy: get really good at counting just a few things, then use bijections
to count everything else! This is the strategy we’ll follow. In particular, we’ll get
really good at counting sequences. When we want to determine the size of some
other set T , we’ll find a bijection from T to a set of sequences S . Then we’ll
use our super-ninja sequence-counting skills to determine jSj, which immediately
gives us jT j. We’ll need to hone this idea somewhat as we go along, but that’s
pretty much it!
“mcs” — 2017/6/5 — 19:42 — page 615 — #623
15.2. Counting Sequences
15.2.1
615
The Product Rule
The Product Rule gives the size of a product of sets. Recall that if P1 ; P2 ; : : : ; Pn
are sets, then
P1 P2 Pn
is the set of all sequences whose first term is drawn from P1 , second term is drawn
from P2 and so forth.
Rule 15.2.1 (Product Rule). If P1 ; P2 ; : : : Pn are finite sets, then:
jP1 P2 Pn j D jP1 j jP2 j jPn j
For example, suppose a daily diet consists of a breakfast selected from set B, a
lunch from set L, and a dinner from set D where:
B D fpancakes; bacon and eggs; bagel; Doritosg
L D fburger and fries; garden salad; Doritosg
D D fmacaroni; pizza; frozen burrito; pasta; Doritosg
Then B LD is the set of all possible daily diets. Here are some sample elements:
.pancakes; burger and fries; pizza/
.bacon and eggs; garden salad; pasta/
.Doritos; Doritos; frozen burrito/
The Product Rule tells us how many different daily diets are possible:
jB L Dj D jBj jLj jDj
D435
D 60:
15.2.2
Subsets of an n-element Set
The fact that there are 2n subsets of an n-element set was proved in Theorem 4.5.5
by setting up a bijection between the subsets and the length-n bit-strings. So the
original problem about subsets was tranformed into a question about sequences—
exactly according to plan! Now we can fill in the missing explanation of why there
are 2n length-n bit-strings: we can write the set of all n-bit sequences as a product
of sets:
f0; 1gn WWD f0; 1g f0; 1g f0; 1g :
ƒ‚
…
„
n terms
Then Product Rule gives the answer:
jf0; 1gn j D jf0; 1gjn D 2n :
“mcs” — 2017/6/5 — 19:42 — page 616 — #624
616
Chapter 15
15.2.3
Cardinality Rules
The Sum Rule
Bart allocates his little sister Lisa a quota of 20 crabby days, 40 irritable days,
and 60 generally surly days. On how many days can Lisa be out-of-sorts one way
or another? Let set C be her crabby days, I be her irritable days, and S be the
generally surly. In these terms, the answer to the question is jC [ I [ S j. Now
assuming that she is permitted at most one bad quality each day, the size of this
union of sets is given by the Sum Rule:
Rule 15.2.2 (Sum Rule). If A1 ; A2 ; : : : ; An are disjoint sets, then:
jA1 [ A2 [ : : : [ An j D jA1 j C jA2 j C : : : C jAn j
Thus, according to Bart’s budget, Lisa can be out-of-sorts for:
jC [ I [ S j D jC j C jI j C jS j
D 20 C 40 C 60
D 120 days
Notice that the Sum Rule holds only for a union of disjoint sets. Finding the size
of a union of overlapping sets is a more complicated problem that we’ll take up in
Section 15.9.
15.2.4
Counting Passwords
Few counting problems can be solved with a single rule. More often, a solution is
a flurry of sums, products, bijections, and other methods.
For solving problems involving passwords, telephone numbers, and license plates,
the sum and product rules are useful together. For example, on a certain computer
system, a valid password is a sequence of between six and eight symbols. The first
symbol must be a letter (which can be lowercase or uppercase), and the remaining symbols must be either letters or digits. How many different passwords are
possible?
Let’s define two sets, corresponding to valid symbols in the first and subsequent
positions in the password.
F D fa; b; : : : ; z; A; B; : : : ; Zg
S D fa; b; : : : ; z; A; B; : : : ; Z; 0; 1; : : : ; 9g
In these terms, the set of all possible passwords is:1
.F S 5 / [ .F S 6 / [ .F S 7 /
1 The
notation S 5 means S S S S S.
“mcs” — 2017/6/5 — 19:42 — page 617 — #625
15.3. The Generalized Product Rule
617
Thus, the length-six passwords are in the set F S 5 , the length-seven passwords
are in F S 6 , and the length-eight passwords are in F S 7 . Since these sets
are disjoint, we can apply the Sum Rule and count the total number of possible
passwords as follows:
j.F S 5 / [ .F S 6 / [ .F S 7 /j
D jF S 5 j C jF S 6 j C jF S 7 j
5
6
7
D jF j jSj C jF j jS j C jF j jS j
Sum Rule
Product Rule
D 52 625 C 52 626 C 52 627
1:8 1014 different passwords:
15.3
The Generalized Product Rule
In how many ways can, say, a Nobel prize, a Japan prize, and a Pulitzer prize be
awarded to n people? This is easy to answer using our strategy of translating the
problem about awards into a problem about sequences. Let P be the set of n people
taking the course. Then there is a bijection from ways of awarding the three prizes
to the set P 3 WWD P P P . In particular, the assignment:
“Barack wins a Nobel, George wins a Japan, and Bill wins a Pulitzer prize”
maps to the sequence .Barack; George; Bill/. By the Product Rule, we have jP 3 j D
jP j3 D n3 , so there are n3 ways to award the prizes to a class of n people. Notice
that P 3 includes triples like .Barack; Bill; Barack/ where one person wins more
than one prize.
But what if the three prizes must be awarded to different students? As before,
we could map the assignment to the triple .Bill; George; Barack/ 2 P 3 . But this
function is no longer a bijection. For example, no valid assignment maps to the
triple .Barack; Bill; Barack/ because now we’re not allowing Barack to receive two
prizes. However, there is a bijection from prize assignments to the set:
S D f.x; y; z/ 2 P 3 j x, y and z are different peopleg
This reduces the original problem to a problem of counting sequences. Unfortunately, the Product Rule does not apply directly to counting sequences of this type
because the entries depend on one another; in particular, they must all be different.
However, a slightly sharper tool does the trick.
“mcs” — 2017/6/5 — 19:42 — page 618 — #626
618
Chapter 15
Cardinality Rules
Prizes for truly exceptional Coursework
Given everyone’s hard work on this material, the instructors considered awarding some prizes for truly exceptional coursework. Here are three possible prize
categories:
Best Administrative Critique We asserted that the quiz was closed-book. On
the cover page, one strong candidate for this award wrote, “There is no
book.”
Awkward Question Award “Okay, the left sock, right sock, and pants are in
an antichain, but how—even with assistance—could I put on all three at
once?”
Best Collaboration Statement Inspired by a student who wrote “I worked alone”
on Quiz 1.
Rule 15.3.1 (Generalized Product Rule). Let S be a set of length-k sequences. If
there are:
n1 possible first entries,
n2 possible second entries for each first entry,
::
:
nk possible kth entries for each sequence of first k
1 entries,
then:
jS j D n1 n2 n3 nk
In the awards example, S consists of sequences .x; y; z/. There are n ways to
choose x, the recipient of prize #1. For each of these, there are n 1 ways to choose
y, the recipient of prize #2, since everyone except for person x is eligible. For each
combination of x and y, there are n 2 ways to choose z, the recipient of prize #3,
because everyone except for x and y is eligible. Thus, according to the Generalized
Product Rule, there are
jS j D n .n 1/ .n 2/
ways to award the 3 prizes to different people.
“mcs” — 2017/6/5 — 19:42 — page 619 — #627
15.3. The Generalized Product Rule
15.3.1
619
Defective Dollar Bills
A dollar bill is defective if some digit appears more than once in the 8-digit serial
number. If you check your wallet, you’ll be sad to discover that defective bills
are all-too-common. In fact, how common are nondefective bills? Assuming that
the digit portions of serial numbers all occur equally often, we could answer this
question by computing
fraction of nondefective bills D
jfserial #’s with all digits differentgj
:
jfserial numbersgj
(15.1)
Let’s first consider the denominator. Here there are no restrictions; there are 10
possible first digits, 10 possible second digits, 10 third digits, and so on. Thus, the
total number of 8-digit serial numbers is 108 by the Product Rule.
Next, let’s turn to the numerator. Now we’re not permitted to use any digit twice.
So there are still 10 possible first digits, but only 9 possible second digits, 8 possible
third digits, and so forth. Thus, by the Generalized Product Rule, there are
10 9 8 7 6 5 4 3 D
10Š
D 1;814;400
2
serial numbers with all digits different. Plugging these results into Equation 15.1,
we find:
fraction of nondefective bills D
15.3.2
1;814;400
D 1:8144%
100;000;000
A Chess Problem
In how many different ways can we place a pawn (P ), a knight (N ), and a bishop
(B) on a chessboard so that no two pieces share a row or a column? A valid configuration is shown in Figure 15.1(a), and an invalid configuration is shown in Figure 15.1(b).
First, we map this problem about chess pieces to a question about sequences.
There is a bijection from configurations to sequences
.rP ; cP ; rN ; cN ; rB ; cB /
where rP , rN and rB are distinct rows and cP , cN and cB are distinct columns.
In particular, rP is the pawn’s row cP is the pawn’s column rN is the knight’s
row, etc. Now we can count the number of such sequences using the Generalized
Product Rule:
rP is one of 8 rows
“mcs” — 2017/6/5 — 19:42 — page 620 — #628
620
Chapter 15
Cardinality Rules
0Z0Z0Z0Z
7
Z0Z0m0Z0
6
0Z0Z0Z0Z
5
Z0Z0Z0Z0
4
0a0Z0Z0Z
3
Z0Z0Z0Z0
2
0Z0Z0o0Z
1
Z0Z0Z0Z0
8
a
b
c
d
e
f
g
h
0Z0Z0Z0Z
Z0Z0Z0Z0
6
0Z0ZpZ0Z
5
Z0Z0Z0Z0
4
0Z0Z0Z0Z
3
Z0a0ZnZ0
2
0Z0Z0Z0Z
1
Z0Z0Z0Z0
8
7
a
b
c
(a) valid
d
e
f
g
h
(b) invalid
Figure 15.1 Two ways of placing a pawn (p), a knight (N), and a bishop (B) on
a chessboard. The configuration shown in (b) is invalid because the bishop and the
knight are in the same row.
cP is one of 8 columns
rN is one of 7 rows (any one but rP )
cN is one of 7 columns (any one but cP )
rB is one of 6 rows (any one but rP or rN )
cB is one of 6 columns (any one but cP or cN )
Thus, the total number of configurations is .8 7 6/2 .
15.3.3
Permutations
A permutation of a set S is a sequence that contains every element of S exactly
once. For example, here are all the permutations of the set fa; b; cg:
.a; b; c/ .a; c; b/ .b; a; c/
.b; c; a/ .c; a; b/ .c; b; a/
How many permutations of an n-element set are there? Well, there are n choices
for the first element. For each of these, there are n 1 remaining choices for the
second element. For every combination of the first two elements, there are n 2
ways to choose the third element, and so forth. Thus, there are a total of
n .n
1/ .n
2/ 3 2 1 D nŠ
permutations of an n-element set. In particular, this formula says that there are
“mcs” — 2017/6/5 — 19:42 — page 621 — #629
15.4. The Division Rule
621
3Š D 6 permutations of the 3-element set fa; b; cg, which is the number we found
above.
Permutations will come up again in this course approximately 1.6 bazillion times.
In fact, permutations are the reason why factorial comes up so often and why we
taught you Stirling’s approximation:
n n
p
nŠ 2 n
:
e
15.4
The Division Rule
Counting ears and dividing by two is a silly way to count the number of people in
a room, but this approach is representative of a powerful counting principle.
A k-to-1 function maps exactly k elements of the domain to every element of
the codomain. For example, the function mapping each ear to its owner is 2-to-1.
Similarly, the function mapping each finger to its owner is 10-to-1, and the function
mapping each finger and toe to its owner is 20-to-1. The general rule is:
Rule 15.4.1 (Division Rule). If f W A ! B is k-to-1, then jAj D k jBj.
For example, suppose A is the set of ears in the room and B is the set of people.
There is a 2-to-1 mapping from ears to people, so by the Division Rule, jAj D
2 jBj. Equivalently, jBj D jAj=2, expressing what we knew all along: the number
of people is half the number of ears. Unlikely as it may seem, many counting
problems are made much easier by initially counting every item multiple times and
then correcting the answer using the Division Rule. Let’s look at some examples.
15.4.1
Another Chess Problem
In how many different ways can you place two identical rooks on a chessboard
so that they do not share a row or column? A valid configuration is shown in
Figure 15.2(a), and an invalid configuration is shown in Figure 15.2(b).
Let A be the set of all sequences
.r1 ; c1 ; r2 ; c2 /
where r1 and r2 are distinct rows and c1 and c2 are distinct columns. Let B be the
set of all valid rook configurations. There is a natural function f from set A to set
B; in particular, f maps the sequence .r1 ; c1 ; r2 ; c2 / to a configuration with one
rook in row r1 , column c1 and the other rook in row r2 , column c2 .
“mcs” — 2017/6/5 — 19:42 — page 622 — #630
622
Chapter 15
Cardinality Rules
0Z0Z0Z0s
7
Z0Z0Z0Z0
6
0Z0Z0Z0Z
5
Z0Z0Z0Z0
4
0Z0Z0Z0Z
3
Z0Z0Z0Z0
2
0Z0Z0Z0Z
1
s0Z0Z0Z0
0Z0Z0Z0Z
Z0Z0Z0Z0
6
0Z0s0Z0Z
5
Z0Z0Z0Z0
4
0Z0Z0Z0Z
3
Z0Z0Z0Z0
2
0Z0Z0Z0Z
1
Z0ZrZ0Z0
8
a
b
c
d
e
f
g
8
7
h
a
b
(a) valid
c
d
e
f
g
h
(b) invalid
Figure 15.2 Two ways to place 2 rooks (R) on a chessboard. The configuration
in (b) is invalid because the rooks are in the same column.
But now there’s a snag. Consider the sequences:
.1; a; 8; h/
and
.8; h; 1; a/
The first sequence maps to a configuration with a rook in the lower-left corner and
a rook in the upper-right corner. The second sequence maps to a configuration with
a rook in the upper-right corner and a rook in the lower-left corner. The problem is
that those are two different ways of describing the same configuration! In fact, this
arrangement is shown in Figure 15.2(a).
More generally, the function f maps exactly two sequences to every board configuration; f is a 2-to-1 function. Thus, by the quotient rule, jAj D 2 jBj. Rearranging terms gives:
jAj
.8 7/2
jBj D
D
:
2
2
In the second equality, we’ve computed the size of A using the General Product
Rule just as in the earlier chess problem.
15.4.2
Knights of the Round Table
In how many ways can King Arthur arrange to seat his n different knights at his
round table? A seating defines who sits where. Two seatings are considered to be
the same arrangement if each knight sits between the same two knights in both
seatings. An equivalent way to say this is that two seatings yield the same arrangement when they yield the same sequence of knights starting at knight number 1
“mcs” — 2017/6/5 — 19:42 — page 623 — #631
15.5. Counting Subsets
623
and going clockwise around the table. For example, the following two seatings
determine the same arrangement:
k
k
1
#
k4
3
#
k2
k2
"!
k4
"!
k3
k1
A seating is determined by the sequence of knights going clockwise around the
table starting at the top seat. So seatings correspond to permutations of the knights,
and there are nŠ of them. For example,
k
2
#
.k2 ; k4 ; k1 ; k3 /
!
k4
k3
"!
k1
Two seatings determine the same arrangement if they are the same when the
table is rotated so knight 1 is at the top seat. For example with n D 4, there are 4
different sequences that correspond to the seating arrangement:
.k2 ; k4 ; k1 ; k3 /
.k4 ; k1 ; k3 ; k2 /
.k1 ; k3 ; k2 ; k4 /
.k3 ; k2 ; k4 ; k1 /
k
1
#
!
k3
k4
"!
k2
This mapping from seating to arrangments is actually an n-to-1 function, since all n
cyclic shifts of the sequence of knights in the seating map to the same arrangement.
Therefore, by the division rule, the number of circular seating arrangements is:
nŠ
# seatings
D
D .n 1/Š :
n
n
15.5
Counting Subsets
How many k-element subsets of an n-element set are there? This question arises
all the time in various guises:
“mcs” — 2017/6/5 — 19:42 — page 624 — #632
624
Chapter 15
Cardinality Rules
In how many ways can I select 5 books from my collection of 100 to bring
on vacation?
How many different 13-card bridge hands can be dealt from a 52-card deck?
In how many ways can I select 5 toppings for my pizza if there are 14 available toppings?
This number comes up so often that there is a special notation for it:
!
n
WWD the number of k-element subsets of an n-element set.
k
The expression kn is read “n choose k.” Now we can immediately express the
answers to all three questions above:
I can select 5 books from 100 in 100
5 ways.
There are 52
13 different bridge hands.
There are 14
5 different 5-topping pizzas, if 14 toppings are available.
15.5.1
The Subset Rule
We can derive a simple formula for the n choose k number using the Division Rule.
We do this by mapping any permutation of an n-element set fa1 ; : : : ; an g into a kelement subset simply by taking the first k elements of the permutation. That is,
the permutation a1 a2 : : : an will map to the set fa1 ; a2 ; : : : ; ak g.
Notice that any other permutation with the same first k elements a1 ; : : : ; ak in
any order and the same remaining elements n k elements in any order will also
map to this set. What’s more, a permutation can only map to fa1 ; a2 ; : : : ; ak g
if its first k elements are the elements a1 ; : : : ; ak in some order. Since there are
kŠ possible permutations of the first k elements and .n k/Š permutations of the
remaining elements, we conclude from the Product Rule that exactly kŠ.n k/Š
permutations of the n-element set map to the particular subset S . In other words,
the mapping from permutations to k-element subsets is kŠ.n k/Š-to-1.
But we know there are nŠ permutations of an n-element set, so by the Division
Rule, we conclude that
!
n
nŠ D kŠ.n k/Š
k
which proves:
“mcs” — 2017/6/5 — 19:42 — page 625 — #633
15.6. Sequences with Repetitions
625
Rule 15.5.1 (Subset Rule). The number of k-element subsets of an n-element set is
!
n
nŠ
:
D
kŠ .n k/Š
k
Notice that this works even for 0-element subsets: nŠ=0ŠnŠ D 1. Here we use the
fact that 0Š is a product of 0 terms, which by convention2 equals 1.
15.5.2
Bit Sequences
How many n-bit sequences contain exactly k ones? We’ve already seen the straightforward bijection between subsets of an n-element set and n-bit sequences. For
example, here is a 3-element subset of fx1 ; x2 ; : : : ; x8 g and the associated 8-bit
sequence:
f x1 ;
x4 ; x5
g
. 1; 0; 0; 1; 1; 0; 0; 0 /
Notice that this sequence has exactly 3 ones, each corresponding to an element
of the 3-element subset. More generally, the n-bit sequences corresponding to a
k-element subset will have exactly k ones. So by the Bijection Rule,
!
n
.
Corollary 15.5.2. The number of n-bit sequences with exactly k ones is
k
Also, the bijection between selections of flavored donuts and bit sequences of
Lemma 15.1.1 now implies,
Corollary 15.5.3. The number of ways to select n donuts when k flavors are available is
!
n C .k 1/
:
n
15.6
Sequences with Repetitions
15.6.1
Sequences of Subsets
Choosing a k-element subset of an n-element set is the same as splitting the set
into a pair of subsets: the first subset of size k and the second subset consisting of
the remaining n k elements. So, the Subset Rule can be understood as a rule for
counting the number of such splits into pairs of subsets.
2 We
don’t use it here, but a sum of zero terms equals 0.
“mcs” — 2017/6/5 — 19:42 — page 626 — #634
626
Chapter 15
Cardinality Rules
We can generalize this to a way to count splits into more than two subsets. Let
A be an n-element set and k1 ; k2 ; : : : ; km be nonnegative integers whose sum is n.
A .k1 ; k2 ; : : : ; km /-split of A is a sequence
.A1 ; A2 ; : : : ; Am /
where the Ai are disjoint subsets of A and jAi j D ki for i D 1; : : : ; m.
To count the number of splits we take the same approach as for the Subset
Rule. Namely, we map any permutation a1 a2 : : : an of an n-element set A into
a .k1 ; k2 ; : : : ; km /-split by letting the 1st subset in the split be the first k1 elements
of the permutation, the 2nd subset of the split be the next k2 elements, . . . , and the
mth subset of the split be the final km elements of the permutation. This map is
a k1 Š k2 Š km Š-to-1 function from the nŠ permutations to the .k1 ; k2 ; : : : ; km /splits of A, so from the Division Rule we conclude the Subset Split Rule:
Definition 15.6.1. For n; k1 ; : : : ; km 2 N, such that k1 Ck2 C Ckm D n, define
the multinomial coefficient
!
n
nŠ
WWD
:
k1 ; k2 ; : : : ; km
k1 Š k2 Š : : : km Š
Rule 15.6.2 (Subset Split Rule). The number of .k1 ; k2 ; : : : ; km /-splits of an nelement set is
!
n
:
k1 ; : : : ; km
15.6.2
The Bookkeeper Rule
We can also generalize our count of n-bit sequences with k ones to counting sequences of n letters over an alphabet with more than two letters. For example,
how many sequences can be formed by permuting the letters in the 10-letter word
BOOKKEEPER?
Notice that there are 1 B, 2 O’s, 2 K’s, 3 E’s, 1 P, and 1 R in BOOKKEEPER. This
leads to a straightforward bijection between permutations of BOOKKEEPER and
(1,2,2,3,1,1)-splits of f1; 2; : : : ; 10g. Namely, map a permutation to the sequence
of sets of positions where each of the different letters occur.
For example, in the permutation BOOKKEEPER itself, the B is in the 1st position, the O’s occur in the 2nd and 3rd positions, K’s in 4th and 5th, the E’s in the
6th, 7th and 9th, P in the 8th, and R is in the 10th position. So BOOKKEEPER
maps to
.f1g; f2; 3g; f4; 5g; f6; 7; 9g; f8g; f10g/:
“mcs” — 2017/6/5 — 19:42 — page 627 — #635
15.6. Sequences with Repetitions
627
From this bijection and the Subset Split Rule, we conclude that the number of ways
to rearrange the letters in the word BOOKKEEPER is:
total letters
‚…„ƒ
10Š
1Š „ƒ‚…
2Š „ƒ‚…
2Š „ƒ‚…
3Š „ƒ‚…
1Š „ƒ‚…
1Š
„ƒ‚…
B’s
O’s
K’s
E’s
P’s
R’s
This example generalizes directly to an exceptionally useful counting principle
which we will call the
Rule 15.6.3 (Bookkeeper Rule). Let l1 ; : : : ; lm be distinct elements. The number
of sequences with k1 occurrences of l1 , and k2 occurrences of l2 , . . . , and km
occurrences of lm is
!
k1 C k2 C C km
:
k1 ; : : : ; km
For example, suppose you are planning a 20-mile walk, which should include 5
northward miles, 5 eastward miles, 5 southward miles, and 5 westward miles. How
many different walks are possible?
There is a bijection between such walks and sequences with 5 N’s, 5 E’s, 5 S’s,
and 5 W’s. By the Bookkeeper Rule, the number of such sequences is:
20Š
:
.5Š/4
A Word about Words
Someday you might refer to the Subset Split Rule or the Bookkeeper Rule in front
of a roomful of colleagues and discover that they’re all staring back at you blankly.
This is not because they’re dumb, but rather because we made up the name “Bookkeeper Rule.” However, the rule is excellent and the name is apt, so we suggest
that you play through: “You know? The Bookkeeper Rule? Don’t you guys know
anything?”
The Bookkeeper Rule is sometimes called the “formula for permutations with
indistinguishable objects.” The size k subsets of an n-element set are sometimes
called k-combinations. Other similar-sounding descriptions are “combinations with
repetition, permutations with repetition, r-permutations, permutations with indistinguishable objects,” and so on. However, the counting rules we’ve taught you are
sufficient to solve all these sorts of problems without knowing this jargon, so we
won’t burden you with it.
“mcs” — 2017/6/5 — 19:42 — page 628 — #636
628
Chapter 15
15.6.3
Cardinality Rules
The Binomial Theorem
Counting gives insight into one of the basic theorems of algebra. A binomial is a
sum of two terms, such as a C b. Now consider its fourth power .a C b/4 .
By repeatedly using distributivity of products over sums to multiply out this 4th
power expression completely, we get
.a C b/4 D
aaaa
C abaa
C baaa
C bbaa
C
C
C
C
aaab
abab
baab
bbab
C
C
C
C
aaba
abba
baba
bbba
C
C
C
C
aabb
abbb
babb
bbbb
Notice that there is one term for every sequence of a’s and b’s. So there are 24
terms, and the number of terms with k copies of b and n k copies of a is:
!
nŠ
n
D
kŠ .n k/Š
k
by the Bookkeeper Rule. Hence, the coefficient of an k b k is kn . So for n D 4,
this means:
!
!
!
!
!
4
4
4
4
4
4
4 0
3 1
2 2
1 3
.a C b/ D
a b C
a b C
a b C
a b C
a0 b 4
0
1
2
3
4
In general, this reasoning gives the Binomial Theorem:
Theorem 15.6.4 (Binomial Theorem). For all n 2 N and a; b 2 R:
!
n
X
n n k k
n
a
b
.a C b/ D
k
kD0
The Binomial Theorem explains why the n choose k number is called a binomial
coefficient.
This reasoning about binomials extends nicely to multinomials, which are sums
of two or more terms. For example, suppose we wanted the coefficient of
bo2 k 2 e 3 pr
in the expansion of .b C o C k C e C p C r/10 . Each term in this expansion is a
product of 10 variables where each variable is one of b, o, k, e, p or r. Now, the
coefficient of bo2 k 2 e 3 pr is the number of those terms with exactly 1 b, 2 o’s, 2
“mcs” — 2017/6/5 — 19:42 — page 629 — #637
15.7. Counting Practice: Poker Hands
629
k’s, 3 e’s, 1 p and 1 r. And the number of such terms is precisely the number of
rearrangements of the word BOOKKEEPER:
!
10
10Š
D
:
1; 2; 2; 3; 1; 1
1Š 2Š 2Š 3Š 1Š 1Š
This reasoning extends to a general theorem:
Theorem 15.6.5 (Multinomial Theorem). For all n 2 N,
.z1 C z2 C C zm /n D
X
k1 ;:::;km 2N
k1 CCkm Dn
!
n
km
z k1 z k2 zm
:
k1 ; k2 ; : : : ; km 1 2
But you’ll be better off remembering the reasoning behind the Multinomial Theorem rather than this cumbersome formal statement.
15.7
Counting Practice: Poker Hands
Five-Card Draw is a card game in which each player is initially dealt a hand consisting of 5 cards from a deck of 52 cards.3 The number of different hands in
Five-Card Draw is the number of 5-element subsets of a 52-element set, which is
!
52
D 2; 598; 960:
5
Let’s get some counting practice by working out the number of hands with various
special properties.
3 There
are 52 cards in a standard deck. Each card has a suit and a rank. There are four suits:
 (spades)
~ (hearts)
| (clubs)
} (diamonds)
And there are 13 ranks, listed here from lowest to highest:
Ace
Jack
Queen
King
A; 2; 3; 4; 5; 6; 7; 8; 9; J ; Q ; K :
Thus, for example, 8~ is the 8 of hearts and A is the ace of spades.
“mcs” — 2017/6/5 — 19:42 — page 630 — #638
630
Chapter 15
15.7.1
Cardinality Rules
Hands with a Four-of-a-Kind
A Four-of-a-Kind is a set of four cards with the same rank. How many different
hands contain a Four-of-a-Kind? Here are a couple examples:
f8; 8}; Q~; 8~; 8|g
fA|; 2|; 2~; 2}; 2g
As usual, the first step is to map this question to a sequence-counting problem. A
hand with a Four-of-a-Kind is completely described by a sequence specifying:
1. The rank of the four cards.
2. The rank of the extra card.
3. The suit of the extra card.
Thus, there is a bijection between hands with a Four-of-a-Kind and sequences consisting of two distinct ranks followed by a suit. For example, the three hands above
are associated with the following sequences:
.8; Q; ~/ $ f 8; 8}; 8~; 8|; Q~g
.2; A; |/ $ f2|; 2~; 2}; 2; A|g
Now we need only count the sequences. There are 13 ways to choose the first rank,
12 ways to choose the second rank, and 4 ways to choose the suit. Thus, by the
Generalized Product Rule, there are 13 12 4 D 624 hands with a Four-of-a-Kind.
This means that only 1 hand in about 4165 has a Four-of-a-Kind. Not surprisingly,
Four-of-a-Kind is considered to be a very good poker hand!
15.7.2
Hands with a Full House
A Full House is a hand with three cards of one rank and two cards of another rank.
Here are some examples:
f2; 2|; 2}; J |; J }g
f5}; 5|; 5~; 7~; 7|g
Again, we shift to a problem about sequences. There is a bijection between Full
Houses and sequences specifying:
“mcs” — 2017/6/5 — 19:42 — page 631 — #639
15.7. Counting Practice: Poker Hands
631
1. The rank of the triple, which can be chosen in 13 ways.
2. The suits of the triple, which can be selected in 43 ways.
3. The rank of the pair, which can be chosen in 12 ways.
4. The suits of the pair, which can be selected in 42 ways.
The example hands correspond to sequences as shown below:
.2; f; |; }g; J; f|; }g/ $ f2; 2|; 2}; J |; J }g
.5; f}; |; ~g; 7; f~; |g/ $ f5}; 5|; 5~; 7~; 7|g
By the Generalized Product Rule, the number of Full Houses is:
!
!
4
4
13 12 :
3
2
We’re on a roll—but we’re about to hit a speed bump.
15.7.3
Hands with Two Pairs
How many hands have Two Pairs; that is, two cards of one rank, two cards of
another rank, and one card of a third rank? Here are examples:
f3}; 3; Q}; Q~; A|g
f9~; 9}; 5~; 5|; Kg
Each hand with Two Pairs is described by a sequence consisting of:
1. The rank of the first pair, which can be chosen in 13 ways.
2. The suits of the first pair, which can be selected 42 ways.
3. The rank of the second pair, which can be chosen in 12 ways.
4. The suits of the second pair, which can be selected in 42 ways.
5. The rank of the extra card, which can be chosen in 11 ways.
6. The suit of the extra card, which can be selected in 41 D 4 ways.
“mcs” — 2017/6/5 — 19:42 — page 632 — #640
632
Chapter 15
Cardinality Rules
Thus, it might appear that the number of hands with Two Pairs is:
!
!
4
4
13 12 11 4:
2
2
Wrong answer! The problem is that there is not a bijection from such sequences to
hands with Two Pairs. This is actually a 2-to-1 mapping. For example, here are the
pairs of sequences that map to the hands given above:
.3; f}; g; Q; f}; ~g; A; |/ &
f3}; 3; Q}; Q~; A|g
.Q; f}; ~g; 3; f}; g; A; |/ %
.9; f~; }g; 5; f~; |g; K; / &
f9~; 9}; 5~; 5|; Kg
.5; f~; |g; 9; f~; }g; K; / %
The problem is that nothing distinguishes the first pair from the second. A pair of
5’s and a pair of 9’s is the same as a pair of 9’s and a pair of 5’s. We avoided this
difficulty in counting Full Houses because, for example, a pair of 6’s and a triple of
kings is different from a pair of kings and a triple of 6’s.
We ran into precisely this difficulty last time, when we went from counting arrangements of different pieces on a chessboard to counting arrangements of two
identical rooks. The solution then was to apply the Division Rule, and we can do
the same here. In this case, the Division rule says there are twice as many sequences
as hands, so the number of hands with Two Pairs is actually:
13 42 12 42 11 4
:
2
Another Approach
The preceding example was disturbing! One could easily overlook the fact that the
mapping was 2-to-1 on an exam, fail the course, and turn to a life of crime. You
can make the world a safer place in two ways:
1. Whenever you use a mapping f W A ! B to translate one counting problem
to another, check that the same number of elements in A are mapped to each
element in B. If k elements of A map to each of element of B, then apply
the Division Rule using the constant k.
2. As an extra check, try solving the same problem in a different way. Multiple
approaches are often available—and all had better give the same answer!
“mcs” — 2017/6/5 — 19:42 — page 633 — #641
15.7. Counting Practice: Poker Hands
633
(Sometimes different approaches give answers that look different, but turn
out to be the same after some algebra.)
We already used the first method; let’s try the second. There is a bijection between hands with two pairs and sequences that specify:
1. The ranks of the two pairs, which can be chosen in 13
2 ways.
2. The suits of the lower-rank pair, which can be selected in 42 ways.
3. The suits of the higher-rank pair, which can be selected in 42 ways.
4. The rank of the extra card, which can be chosen in 11 ways.
5. The suit of the extra card, which can be selected in 41 D 4 ways.
For example, the following sequences and hands correspond:
.f3; Qg; f}; g; f}; ~g; A; |/ $ f3}; 3; Q}; Q~; A|g
.f9; 5g; f~; |g; f~; }g; K; / $ f9~; 9}; 5~; 5|; Kg
Thus, the number of hands with two pairs is:
!
!
!
13
4
4
11 4:
2
2
2
This is the same answer we got before, though in a slightly different form.
15.7.4
Hands with Every Suit
How many hands contain at least one card from every suit? Here is an example of
such a hand:
f7}; K|; 3}; A~; 2g
Each such hand is described by a sequence that specifies:
1. The ranks of the diamond, the club, the heart, and the spade, which can be
selected in 13 13 13 13 D 134 ways.
2. The suit of the extra card, which can be selected in 4 ways.
3. The rank of the extra card, which can be selected in 12 ways.
“mcs” — 2017/6/5 — 19:42 — page 634 — #642
634
Chapter 15
Cardinality Rules
For example, the hand above is described by the sequence:
.7; K; A; 2; }; 3/ $ f7}; K|; A~; 2; 3}g:
Are there other sequences that correspond to the same hand? There is one more!
We could equally well regard either the 3} or the 7} as the extra card, so this
is actually a 2-to-1 mapping. Here are the two sequences corresponding to the
example hand:
.7; K; A; 2; }; 3/ &
f7}; K|; A~; 2; 3}g
.3; K; A; 2; }; 7/ %
Therefore, the number of hands with every suit is:
134 4 12
:
2
15.8
The Pigeonhole Principle
Here is an old puzzle:
A drawer in a dark room contains red socks, green socks, and blue
socks. How many socks must you withdraw to be sure that you have a
matching pair?
For example, picking out three socks is not enough; you might end up with one
red, one green, and one blue. The solution relies on the
Pigeonhole Principle
If there are more pigeons than holes they occupy, then at least two
pigeons must be in the same hole.
What pigeons have to do with selecting footwear under poor lighting conditions
may not be immediately obvious, but if we let socks be pigeons and the colors be
three pigeonholes, then as soon as you pick four socks, there are bound to be two
in the same hole, that is, with the same color. So four socks are enough to ensure
a matched pair. For example, one possible mapping of four socks to three colors is
shown in Figure 15.3.
A rigorous statement of the Principle goes this way:
“mcs” — 2017/6/5 — 19:42 — page 635 — #643
15.8. The Pigeonhole Principle
A
635
f
B
1st sock
red
2nd sock
green
3rd sock
blue
4th sock
Figure 15.3 One possible mapping of four socks to three colors.
Rule 15.8.1 (Pigeonhole Principle). If jAj > jBj, then for every total function
f W A ! B, there exist two different elements of A that are mapped by f to the
same element of B.
Stating the Principle this way may be less intuitive, but it should now sound
familiar: it is simply the contrapositive of the Mapping Rules injective case (4.6).
Here, the pigeons form set A, the pigeonholes are the set B, and f describes which
hole each pigeon occupies.
Mathematicians have come up with many ingenious applications for the pigeonhole principle. If there were a cookbook procedure for generating such arguments,
we’d give it to you. Unfortunately, there isn’t one. One helpful tip, though: when
you try to solve a problem with the pigeonhole principle, the key is to clearly identify three things:
1. The set A (the pigeons).
2. The set B (the pigeonholes).
3. The function f (the rule for assigning pigeons to pigeonholes).
15.8.1
Hairs on Heads
There are a number of generalizations of the pigeonhole principle. For example:
Rule 15.8.2 (Generalized Pigeonhole Principle). If jAj > k jBj, then every total
function f W A ! B maps at least k C1 different elements of A to the same element
of B.
“mcs” — 2017/6/5 — 19:42 — page 636 — #644
636
Chapter 15
Cardinality Rules
For example, if you pick two people at random, surely they are extremely unlikely to have exactly the same number of hairs on their heads. However, in the
remarkable city of Boston, Massachusetts, there is a group of three people who
have exactly the same number of hairs! Of course, there are many completely bald
people in Boston, and they all have zero hairs. But we’re talking about non-bald
people; say a person is non-bald if they have at least ten thousand hairs on their
head.
Boston has about 500,000 non-bald people, and the number of hairs on a person’s
head is at most 200,000. Let A be the set of non-bald people in Boston, let B D
f10; 000; 10; 001; : : : ; 200; 000g, and let f map a person to the number of hairs on
his or her head. Since jAj > 2jBj, the Generalized Pigeonhole Principle implies
that at least three people have exactly the same number of hairs. We don’t know
who they are, but we know they exist!
15.8.2
Subsets with the Same Sum
For your reading pleasure, we have displayed ninety 25-digit numbers in Figure 15.4. Are there two different subsets of these 25-digit numbers that have the
same sum? For example, maybe the sum of the last ten numbers in the first column
is equal to the sum of the first eleven numbers in the second column?
Finding two subsets with the same sum may seem like a silly puzzle, but solving
these sorts of problems turns out to be useful in diverse applications such as finding
good ways to fit packages into shipping containers and decoding secret messages.
It turns out that it is hard to find different subsets with the same sum, which
is why this problem arises in cryptography. But it is easy to prove that two such
subsets exist. That’s where the Pigeonhole Principle comes in.
Let A be the collection of all subsets of the 90 numbers in the list. Now the
sum of any subset of numbers is at most 90 1025 , since there are only 90 numbers
and every 25-digit number is less than 1025 . So let B be the the integer interval
Zi nt v090 1025 , and let f map each subset of numbers (in A) to its sum (in B).
We proved that an n-element set has 2n different subsets in Section 15.2. Therefore:
jAj D 290 1:237 1027
On the other hand:
jBj D 90 1025 C 1 0:901 1027 :
Both quantities are enormous, but jAj is a bit greater than jBj. This means that f
maps at least two elements of A to the same element of B. In other words, by the
Pigeonhole Principle, two different subsets must have the same sum!
“mcs” — 2017/6/5 — 19:42 — page 637 — #645
15.8. The Pigeonhole Principle
0020480135385502964448038
5763257331083479647409398
0489445991866915676240992
5800949123548989122628663
1082662032430379651370981
6042900801199280218026001
1178480894769706178994993
6116171789137737896701405
1253127351683239693851327
6144868973001582369723512
1301505129234077811069011
6247314593851169234746152
1311567111143866433882194
6814428944266874963488274
1470029452721203587686214
6870852945543886849147881
1578271047286257499433886
6914955508120950093732397
1638243921852176243192354
6949632451365987152423541
1763580219131985963102365
7128211143613619828415650
1826227795601842231029694
7173920083651862307925394
1843971862675102037201420
7215654874211755676220587
2396951193722134526177237
7256932847164391040233050
2781394568268599801096354
7332822657075235431620317
2796605196713610405408019
7426441829541573444964139
2931016394761975263190347
7632198126531809327186321
2933458058294405155197296
7712154432211912882310511
3075514410490975920315348
7858918664240262356610010
8149436716871371161932035
3111474985252793452860017
7898156786763212963178679
3145621587936120118438701
8147591017037573337848616
3148901255628881103198549
5692168374637019617423712
637
3171004832173501394113017
8247331000042995311646021
3208234421597368647019265
8496243997123475922766310
3437254656355157864869113
8518399140676002660747477
3574883393058653923711365
8543691283470191452333763
3644909946040480189969149
8675309258374137092461352
3790044132737084094417246
8694321112363996867296665
3870332127437971355322815
8772321203608477245851154
4080505804577801451363100
8791422161722582546341091
4167283461025702348124920
9062628024592126283973285
4235996831123777788211249
9137845566925526349897794
4670939445749439042111220
9153762966803189291934419
4815379351865384279613427
9270880194077636406984249
4837052948212922604442190
9324301480722103490379204
5106389423855018550671530
9436090832146695147140581
5142368192004769218069910
9475308159734538249013238
5181234096130144084041856
9492376623917486974923202
5198267398125617994391348
9511972558779880288252979
5317592940316231219758372
9602413424619187112552264
5384358126771794128356947
9631217114906129219461111
3157693105325111284321993
5439211712248901995423441
9908189853102753335981319
5610379826092838192760458
9913237476341764299813987
5632317555465228677676044
8176063831682536571306791
Figure 15.4 Ninety 25-digit numbers. Can you find two different subsets of these
numbers that have the same sum?
“mcs” — 2017/6/5 — 19:42 — page 638 — #646
638
Chapter 15
Cardinality Rules
Notice that this proof gives no indication which two sets of numbers have the
same sum. This frustrating variety of argument is called a nonconstructive proof.
The $100 prize for two same-sum subsets
To see if it was possible to actually find two different subsets of the ninety 25-digit
numbers with the same sum, we offered a $100 prize to the first student who did it.
We didn’t expect to have to pay off this bet, but we underestimated the ingenuity
and initiative of the students. One computer science major wrote a program that
cleverly searched only among a reasonably small set of “plausible” sets, sorted
them by their sums, and actually found a couple with the same sum. He won the
prize. A few days later, a math major figured out how to reformulate the sum
problem as a “lattice basis reduction” problem; then he found a software package
implementing an efficient basis reduction procedure, and using it, he very quickly
found lots of pairs of subsets with the same sum. He didn’t win the prize, but he
got a standing ovation from the class—staff included.
The $500 Prize for Sets with Distinct Subset Sums
How can we construct a set of n positive integers such that all its subsets have
distinct sums? One way is to use powers of two:
f1; 2; 4; 8; 16g
This approach is so natural that one suspects all other such sets must involve
larger numbers. (For example, we could safely replace 16 by 17, but not by 15.)
Remarkably, there are examples involving smaller numbers. Here is one:
f6; 9; 11; 12; 13g
One of the top mathematicians of the Twentieth Century, Paul Erdős, conjectured
in 1931 that there are no such sets involving significantly smaller numbers. More
precisely, he conjectured that the largest number in such a set must be greater
than c2n for some constant c > 0. He offered $500 to anyone who could prove
or disprove his conjecture, but the problem remains unsolved.
15.8.3
A Magic Trick
A Magician sends an Assistant into the audience with a deck of 52 cards while the
Magician looks away.
Five audience members each select one card from the deck. The Assistant then
gathers up the five cards and holds up four of them so the Magician can see them.
“mcs” — 2017/6/5 — 19:42 — page 639 — #647
15.8. The Pigeonhole Principle
639
The Magician concentrates for a short time and then correctly names the secret,
fifth card!
Since we don’t really believe the Magician can read minds, we know the Assistant has somehow communicated the secret card to the Magician. Real Magicians
and Assistants are not to be trusted, so we expect that the Assistant would secretly
signal the Magician with coded phrases or body language, but for this trick they
don’t have to cheat. In fact, the Magician and Assistant could be kept out of sight
of each other while some audience member holds up the 4 cards designated by the
Assistant for the Magician to see.
Of course, without cheating, there is still an obvious way the Assistant can communicate to the Magician: he can choose any of the 4Š D 24 permutations of the
4 cards as the order in which to hold up the cards. However, this alone won’t
quite work: there are 48 cards remaining in the deck, so the Assistant doesn’t have
enough choices of orders to indicate exactly what the secret card is (though he
could narrow it down to two cards).
15.8.4
The Secret
The method the Assistant can use to communicate the fifth card exactly is a nice
application of what we know about counting and matching.
The Assistant has a second legitimate way to communicate: he can choose which
of the five cards to keep hidden. Of course, it’s not clear how the Magician could
determine which of these five possibilities the Assistant selected by looking at the
four visible cards, but there is a way, as we’ll now explain.
The problem facing the Magician and Assistant is actually a bipartite matching
problem. Each vertex on the left will correspond to the information available to the
Assistant, namely, a set of 5 cards. So the set X of left-hand vertices will have 52
5
elements.
Each vertex on the right will correspond to the information available to the Magician, namely, a sequence of 4 distinct cards. So the set Y of right-hand vertices
will have 52 51 50 49 elements. When the audience selects a set of 5 cards, then
the Assistant must reveal a sequence of 4 cards from that hand. This constraint is
represented by having an edge between a set of 5 cards on the left and a sequence
of 4 cards on the right precisely when every card in the sequence is also in the
set. This specifies the bipartite graph. Some edges are shown in the diagram in
Figure 15.5.
For example,
f8~; K; Q; 2}; 6}g
(15.2)
is an element of X on the left. If the audience selects this set of 5 cards, then
there are many different 4-card sequences on the right in set Y that the Assis-
“mcs” — 2017/6/5 — 19:42 — page 640 — #648
640
Chapter 15
Cardinality Rules
yDall
sequences of 4
distinct cards



xDall
sets of
5 cards



f8~;K;Q;2}g
fK;8~;Q;2}g
f8~;K;Q;2};6}g



f8~;K;Q;9|;6}g



fK;8~;6};Qg



Figure 15.5 The bipartite graph where the nodes on the left correspond to sets
of 5 cards and the nodes on the right correspond to sequences of 4 cards. There is
an edge between a set and a sequence whenever all the cards in the sequence are
contained in the set.
tant could choose to reveal, including .8~; K; Q; 2}/, .K; 8~; Q; 2}/ and
.K; 8~; 6}; Q/.
What the Magician and his Assistant need to perform the trick is a matching for
the X vertices. If they agree in advance on some matching, then when the audience
selects a set of 5 cards, the Assistant reveals the matching sequence of 4 cards. The
Magician uses the matching to find the audience’s chosen set of 5 cards, and so he
can name the one not already revealed.
For example, suppose the Assistant and Magician agree on a matching containing
the two bold edges in Figure 15.5. If the audience selects the set
f8~; K; Q; 9|; 6}g;
(15.3)
then the Assistant reveals the corresponding sequence
.K; 8~; 6}; Q/:
(15.4)
Using the matching, the Magician sees that the hand (15.3) is matched to the sequence (15.4), so he can name the one card in the corresponding set not already
revealed, namely, the 9|. Notice that the fact that the sets are matched, that is,
that different sets are paired with distinct sequences, is essential. For example, if
the audience picked the previous hand (15.2), it would be possible for the Assistant
“mcs” — 2017/6/5 — 19:42 — page 641 — #649
15.8. The Pigeonhole Principle
641
to reveal the same sequence (15.4), but he better not do that; if he did, then the
Magician would have no way to tell if the remaining card was the 9| or the 2}.
So how can we be sure the needed matching can be found? The answer is that
each vertex on the left has degree 54Š D 120, since there are five ways to select the
card kept secret and there are 4Š permutations of the remaining 4 cards. In addition,
each vertex on the right has degree 48, since there are 48 possibilities for the fifth
card. So this graph is degree-constrained according to Definition 12.5.5, and so has
a matching by Theorem 12.5.6.
In fact, this reasoning shows that the Magician could still pull off the trick if 120
cards were left instead of 48, that is, the trick would work with a deck as large as
124 different cards—without any magic!
15.8.5
The Real Secret
But wait a minute! It’s all very well in principle to have the Magician and his
Assistant agree on a matching, but how are they supposed to remember a matching
with 52
5 D 2; 598; 960 edges? For the trick to work in practice, there has to be a
way to match hands and card sequences mentally and on the fly.
We’ll describe one approach. As a running example, suppose that the audience
selects:
10~ 9} 3~ Q J }:
The Assistant picks out two cards of the same suit. In the example, the
assistant might choose the 3~ and 10~. This is always possible because of
the Pigeonhole Principle—there are five cards and 4 suits so two cards must
be in the same suit.
The Assistant locates the ranks of these two cards on the cycle shown in Figure 15.6. For any two distinct ranks on this cycle, one is always between 1
and 6 hops clockwise from the other. For example, the 3~ is 6 hops clockwise from the 10~.
The more counterclockwise of these two cards is revealed first, and the other
becomes the secret card. Thus, in our example, the 10~ would be revealed,
and the 3~ would be the secret card. Therefore:
– The suit of the secret card is the same as the suit of the first card revealed.
– The rank of the secret card is between 1 and 6 hops clockwise from the
rank of the first card revealed.
“mcs” — 2017/6/5 — 19:42 — page 642 — #650
642
Chapter 15
Cardinality Rules
A
K
2
3
Q
J
4
5
10
9
6
8
7
Figure 15.6 The 13 card ranks arranged in cyclic order.
All that remains is to communicate a number between 1 and 6. The Magician
and Assistant agree beforehand on an ordering of all the cards in the deck
from smallest to largest such as:
A| A} A~ A 2| 2} 2~ 2 : : : K~ K
The order in which the last three cards are revealed communicates the number according to the following scheme:
. small; medium;
large / = 1
. small;
large;
medium / = 2
. medium; small;
large / = 3
. medium;
large;
small / = 4
. large;
small; medium / = 5
. large;
medium; small / = 6
In the example, the Assistant wants to send 6 and so reveals the remaining
three cards in large, medium, small order. Here is the complete sequence that
the Magician sees:
10~ Q J } 9}
The Magician starts with the first card 10~ and hops 6 ranks clockwise to
reach 3~, which is the secret card!
So that’s how the trick can work with a standard deck of 52 cards. On the other
hand, Hall’s Theorem implies that the Magician and Assistant can in principle perform the trick with a deck of up to 124 cards. It turns out that there is a method
“mcs” — 2017/6/5 — 19:42 — page 643 — #651
15.9. Inclusion-Exclusion
643
which they could actually learn to use with a reasonable amount of practice for a
124-card deck, but we won’t explain it here.4
15.8.6
The Same Trick with Four Cards?
Suppose that the audience selects only four cards and the Assistant reveals a sequence of three to the Magician. Can the Magician determine the fourth card?
Let X be all the sets of four cards that the audience might select, and let Y be all
the sequences of three cards that the Assistant might reveal. Now, on one hand, we
have
!
52
jXj D
D 270; 725
4
by the Subset Rule. On the other hand, we have
jY j D 52 51 50 D 132; 600
by the Generalized Product Rule. Thus, by the Pigeonhole Principle, the Assistant
must reveal the same sequence of three cards for at least
270; 725
D3
132; 600
different four-card hands. This is bad news for the Magician: if he sees that sequence of three, then there are at least three possibilities for the fourth card which
he cannot distinguish. So there is no legitimate way for the Assistant to communicate exactly what the fourth card is!
15.9
Inclusion-Exclusion
How big is a union of sets? For example, suppose there are 60 math majors, 200
EECS majors, and 40 physics majors. How many students are there in these three
departments? Let M be the set of math majors, E be the set of EECS majors, and
P be the set of physics majors. In these terms, we’re asking for jM [ E [ P j.
The Sum Rule says that if M , E and P are disjoint, then the sum of their sizes
is
jM [ E [ P j D jM j C jEj C jP j:
However, the sets M , E and P might not be disjoint. For example, there might be a
student majoring in both math and physics. Such a student would be counted twice
4 See
The Best Card Trick by Michael Kleber for more information.
“mcs” — 2017/6/5 — 19:42 — page 644 — #652
644
Chapter 15
Cardinality Rules
on the right side of this equation, once as an element of M and once as an element
of P . Worse, there might be a triple-major5 counted three times on the right side!
Our most-complicated counting rule determines the size of a union of sets that
are not necessarily disjoint. Before we state the rule, let’s build some intuition by
considering some easier special cases: unions of just two or three sets.
15.9.1
Union of Two Sets
For two sets, S1 and S2 , the Inclusion-Exclusion Rule is that the size of their union
is:
jS1 [ S2 j D jS1 j C jS2 j jS1 \ S2 j
(15.5)
Intuitively, each element of S1 is accounted for in the first term, and each element
of S2 is accounted for in the second term. Elements in both S1 and S2 are counted
twice—once in the first term and once in the second. This double-counting is corrected by the final term.
15.9.2
Union of Three Sets
So how many students are there in the math, EECS, and physics departments? In
other words, what is jM [ E [ P j if:
jM j D 60
jEj D 200
jP j D 40:
The size of a union of three sets is given by a more complicated Inclusion-Exclusion
formula:
jS1 [ S2 [ S3 j D jS1 j C jS2 j C jS3 j
jS1 \ S2 j
jS1 \ S3 j
jS2 \ S3 j
C jS1 \ S2 \ S3 j:
Remarkably, the expression on the right accounts for each element in the union of
S1 , S2 and S3 exactly once. For example, suppose that x is an element of all three
sets. Then x is counted three times (by the jS1 j, jS2 j and jS3 j terms), subtracted
off three times (by the jS1 \ S2 j, jS1 \ S3 j and jS2 \ S3 j terms), and then counted
once more (by the jS1 \ S2 \ S3 j term). The net effect is that x is counted just
once.
5 . . . though
not at MIT anymore.
“mcs” — 2017/6/5 — 19:42 — page 645 — #653
15.9. Inclusion-Exclusion
645
If x is in two sets (say, S1 and S2 ), then x is counted twice (by the jS1 j and
jS2 j terms) and subtracted once (by the jS1 \ S2 j term). In this case, x does not
contribute to any of the other terms, since x … S3 .
So we can’t answer the original question without knowing the sizes of the various
intersections. Let’s suppose that there are:
4
3
11
2
math - EECS double majors
math - physics double majors
EECS - physics double majors
triple majors
Then jM \Ej D 4C2, jM \P j D 3C2, jE \P j D 11C2, and jM \E \P j D 2.
Plugging all this into the formula gives:
jM [ E [ P j D jM j C jEj C jP j
jM \ Ej
jM \ P j
jE \ P j
C jM \ E \ P j
D 60 C 200 C 40
6
5
13 C 2
D 278
15.9.3
Sequences with 42, 04, or 60
In how many permutations of the set f0; 1; 2; : : : ; 9g do either 4 and 2, 0 and 4, or
6 and 0 appear consecutively? For example, none of these pairs appears in:
.7; 2; 9; 5; 4; 1; 3; 8; 0; 6/:
The 06 at the end doesn’t count; we need 60. On the other hand, both 04 and 60
appear consecutively in this permutation:
.7; 2; 5; 6; 0; 4; 3; 8; 1; 9/:
Let P42 be the set of all permutations in which 42 appears. Define P60 and P04
similarly. Thus, for example, the permutation above is contained in both P60 and
P04 , but not P42 . In these terms, we’re looking for the size of the set P42 [ P04 [
P60 .
First, we must determine the sizes of the individual sets, such as P60 . We can use
a trick: group the 6 and 0 together as a single symbol. Then there is an immediate
bijection between permutations of f0; 1; 2; : : : 9g containing 6 and 0 consecutively
and permutations of:
f60; 1; 2; 3; 4; 5; 7; 8; 9g:
For example, the following two sequences correspond:
.7; 2; 5; 6; 0; 4; 3; 8; 1; 9/
! .7; 2; 5; 60; 4; 3; 8; 1; 9/:
“mcs” — 2017/6/5 — 19:42 — page 646 — #654
646
Chapter 15
Cardinality Rules
There are 9Š permutations of the set containing 60, so jP60 j D 9Š by the Bijection
Rule. Similarly, jP04 j D jP42 j D 9Š as well.
Next, we must determine the sizes of the two-way intersections, such as P42 \
P60 . Using the grouping trick again, there is a bijection with permutations of the
set:
f42; 60; 1; 3; 5; 7; 8; 9g:
Thus, jP42 \ P60 j D 8Š. Similarly, jP60 \ P04 j D 8Š by a bijection with the set:
f604; 1; 2; 3; 5; 7; 8; 9g:
And jP42 \ P04 j D 8Š as well by a similar argument. Finally, note that jP60 \
P04 \ P42 j D 7Š by a bijection with the set:
f6042; 1; 3; 5; 7; 8; 9g:
Plugging all this into the formula gives:
jP42 [ P04 [ P60 j D 9Š C 9Š C 9Š
15.9.4
8Š
8Š
8Š C 7Š:
Union of n Sets
The size of a union of n sets is given by the following rule.
Rule 15.9.1 (Inclusion-Exclusion).
jS1 [ S2 [ [ Sn j D
minus
plus
minus
plus
the sum of the sizes of the individual sets
the sizes of all two-way intersections
the sizes of all three-way intersections
the sizes of all four-way intersections
the sizes of all five-way intersections, etc.
The formulas for unions of two and three sets are special cases of this general
rule.
This way of expressing Inclusion-Exclusion is easy to understand and nearly
as precise as expressing it in mathematical symbols, but we’ll need the symbolic
version below, so let’s work on deciphering it now.
We already have a concise notation for the sum of sizes of the individual sets,
namely,
n
X
jSi j:
i D1
“mcs” — 2017/6/5 — 19:42 — page 647 — #655
15.9. Inclusion-Exclusion
647
A “two-way intersection” is a set of the form Si \ Sj for i ¤ j . We regard Sj \ Si
as the same two-way intersection as Si \ Sj , so we can assume that i < j . Now
we can express the sum of the sizes of the two-way intersections as
X
jSi \ Sj j:
1i <j n
Similarly, the sum of the sizes of the three-way intersections is
X
jSi \ Sj \ Sk j:
1i <j <kn
These sums have alternating signs in the Inclusion-Exclusion formula, with the
sum of the k-way intersections getting the sign . 1/k 1 . This finally leads to a
symbolic version of the rule:
Rule (Inclusion-Exclusion).
ˇ n
ˇ
n
ˇ[ ˇ X
ˇ
ˇ
S
D
jSi j
ˇ
iˇ
ˇ
ˇ
iD1
i D1
X
jSi \ Sj j
1i <j n
C
X
jSi \ Sj \ Sk j C 1i <j <kn
C . 1/n
ˇ n
ˇ
ˇ\ ˇ
ˇ
ˇ
1
Si ˇ :
ˇ
ˇ
ˇ
i D1
While it’s often handy express the rule in this way as a sum of sums, it is not
necessary to group the terms by how many sets are in the intersections. So another
way to state the rule is:
Rule (Inclusion-Exclusion-II).
ˇ n
ˇ
ˇ[ ˇ
ˇ
ˇ
Si ˇ D
ˇ
ˇ
ˇ
i D1
X
;¤I f1;:::;ng
ˇ
ˇ
ˇ\ ˇ
ˇ
ˇ
. 1/jI jC1 ˇ Si ˇ
ˇ
ˇ
(15.6)
i 2I
A proof of these rules using just highschool algebra is given in Problem 15.58.
“mcs” — 2017/6/5 — 19:42 — page 648 — #656
648
Chapter 15
15.9.5
Cardinality Rules
Computing Euler’s Function
We can also use Inclusion-Exclusion to derive the explicit formula for Euler’s funcem
tion claimed in Corollary 9.10.11: if the prime factorization of n is p1e1 pm
for
distinct primes pi , then
m Y
.n/ D n
1
i D1
1
pi
:
(15.7)
To begin, let S be the set of integers in Œ0::n/ that are not relatively prime to n.
So .n/ D n jS j. Next, let Ca be the set of integers in Œ0::n/ that are divisible by
a:
Ca WWD fk 2 Œ0::n/ j a j kg:
So the integers in S are precisely the integers in Œ0::n/ that are divisible by at least
one of the pi ’s. Namely,
m
[
SD
Cpi :
(15.8)
i D1
We’ll be able to find the size of this union using Inclusion-Exclusion because the
intersections of the Cpi ’s are easy to count. For example, Cp \ Cq \ Cr is the set
of integers in Œ0::n/ that are divisible by each of p, q and r. But since the p; q; r
are distinct primes, being divisible by each of them is the same as being divisible
by their product. Now if k is a positive divisor of n, then there are exactly n=k
multiples of k in Œ0::n/. So exactly n=pqr of the integers in Œ0::n/ are divisible by
all three primes p, q, r. In other words,
jCp \ Cq \ Cr j D
n
:
pqr
This reasoning extends to arbitrary intersections of Cp ’s, namely,
ˇ
ˇ
ˇ\
ˇ
ˇ
ˇ
n
ˇ
Cpj ˇˇ D Q
;
ˇ
ˇ
ˇ
j 2I pj
j 2I
(15.9)
“mcs” — 2017/6/5 — 19:42 — page 649 — #657
15.10. Combinatorial Proofs
649
for any nonempty set I Œ1::m. This lets us calculate:
ˇm
ˇ
ˇ[
ˇ
ˇ
ˇ
jS j D ˇ
Cpi ˇ
(by (15.8))
ˇ
ˇ
iD1
ˇ
ˇ
ˇ\
ˇ
X
ˇ
ˇ
D
. 1/jI jC1 ˇ Cpi ˇ
(by Inclusion-Exclusion (15.6))
ˇ
ˇ
i 2I
;¤I Œ1::m
X
n
(by (15.9))
D
. 1/jI jC1 Q
j 2I pj
;¤I Œ1::m
D
n
X
;¤I Œ1::m
D
1
j 2I . pj /
!
Q
m Y
n
1
i D1
1
pi
so
.n/ D n
C n;
m Y
jSj D n
1
i D1
1
pi
;
which proves (15.7).
Yikes! That was pretty hairy. Are you getting tired of all that nasty algebra? If
so, then good news is on the way. In the next section, we will show you how to
prove some heavy-duty formulas without using any algebra at all. Just a few words
and you are done. No kidding.
15.10
Combinatorial Proofs
Suppose you have n different T-shirts, but only want to keep k. You could equally
well select the k shirts you want to keep or select the complementary set of n k
shirts you want to throw out. Thus, the number of ways to select k shirts from
among n must be equal to the number of ways to select n k shirts from among n.
Therefore:
!
!
n
n
D
:
k
n k
This is easy to prove algebraically, since both sides are equal to:
nŠ
:
kŠ .n k/Š
“mcs” — 2017/6/5 — 19:42 — page 650 — #658
650
Chapter 15
Cardinality Rules
But we didn’t really have to resort to algebra; we just used counting principles.
Hmmm.. . .
15.10.1
Pascal’s Triangle Identity
Bob, famed Math for Computer Science Teaching Assistant, has decided to try out
for the US Olympic boxing team. After all, he’s watched all of the Rocky movies
and spent hours in front of a mirror sneering, “Yo, you wanna piece a’ me?!” Bob
figures that n people (including himself) are competing for spots on the team and
only k will be selected. As part of maneuvering for a spot on the team, he needs to
work out how many different teams are possible. There are two cases to consider:
Bob is selected for the team, and his k 1 teammates are selected from
among the other n 1 competitors. The number of different teams that can
be formed in this way is:
!
n 1
:
k 1
Bob is not selected for the team, and all k team members are selected from
among the other n 1 competitors. The number of teams that can be formed
this way is:
!
n 1
:
k
All teams of the first type contain Bob, and no team of the second type does;
therefore, the two sets of teams are disjoint. Thus, by the Sum Rule, the total
number of possible Olympic boxing teams is:
!
!
n 1
n 1
C
:
k 1
k
Ted, equally-famed Teaching Assistant, thinks Bob isn’t so tough and so he
might as well also try out. He reasons that n people (including himself) are trying out for k spots. Thus, the number of ways to select the team is simply:
!
n
:
k
Ted and Bob each correctly counted the number of possible boxing teams. Thus,
their answers must be equal. So we know:
“mcs” — 2017/6/5 — 19:42 — page 651 — #659
15.10. Combinatorial Proofs
651
Lemma 15.10.1 (Pascal’s Triangle Identity).
!
!
!
n
n 1
n 1
D
C
:
k
k 1
k
(15.10)
We proved Pascal’s Triangle Identity without any algebra! Instead, we relied
purely on counting techniques.
15.10.2
Giving a Combinatorial Proof
A combinatorial proof is an argument that establishes an algebraic fact by relying
on counting principles. Many such proofs follow the same basic outline:
1. Define a set S .
2. Show that jSj D n by counting one way.
3. Show that jSj D m by counting another way.
4. Conclude that n D m.
In the preceding example, S was the set of all possible Olympic boxing teams. Bob
computed
!
!
n 1
n 1
C
jSj D
k 1
k
by counting one way, and Ted computed
n
jS j D
k
!
by counting another way. Equating these two expressions gave Pascal’s Identity.
Checking a Combinatorial Proof
Combinatorial proofs are based on counting the same thing in different ways. This
is fine when you’ve become practiced at different counting methods, but when in
doubt, you can fall back on bijections and sequence counting to check such proofs.
For example, let’s take a closer look at the combinatorial proof of Pascal’s Identity (15.10). In this case, the set S of things to be counted is the collection of all
size-k subsets of integers in the interval Œ1::n.
“mcs” — 2017/6/5 — 19:42 — page 652 — #660
652
Chapter 15
Cardinality Rules
Now we’ve
already counted S one way, via the Bookkeeper Rule, and found
n
jS j D k . The other “way” corresponds to defining a bijection between S and the
disjoint union of two sets A and B where,
A WWD f.1; X / j X Œ2; n AND jX j D k
1g
B WWD f.0; Y / j Y Œ2; n AND jY j D kg:
Clearly A and B are disjoint since the pairs in the two sets have different first
coordinates, so jA [ Bj D jAj C jBj. Also,
!
n 1
jAj D # specified sets X D
;
k 1
!
n 1
jBj D # specified sets Y D
:
k
Now finding a bijection f W .A [ B/ ! S will prove the identity (15.10). In
particular, we can define
(
X [ f1g if c D .1; X /;
f .c/ WWD
Y
if c D .0; Y /:
It should be obvious that f is a bijection.
15.10.3
A Colorful Combinatorial Proof
The set that gets counted in a combinatorial proof in different ways is usually defined in terms of simple sequences or sets rather than an elaborate story about
Teaching Assistants. Here is another colorful example of a combinatorial argument.
Theorem 15.10.2.
!
!
!
n
X
n
2n
3n
D
r
n r
n
rD0
Proof. We give a combinatorial proof. Let S be all n-card hands that can be dealt
from a deck containing n different red cards and 2n different black cards. First,
note that every 3n-element set has
!
3n
jS j D
n
“mcs” — 2017/6/5 — 19:42 — page 653 — #661
15.11. References
653
n-element subsets.
From another perspective, the number of hands with exactly r red cards is
!
!
n
2n
r
n r
since there are nr ways to choose the r red cards and n2nr ways to choose the
n r black cards. Since the number of red cards can be anywhere from 0 to n, the
total number of n-card hands is:
!
!
n
X
n
2n
jS j D
:
r
n r
rD0
Equating these two expressions for jS j proves the theorem.
Finding a Combinatorial Proof
Combinatorial proofs are almost magical. Theorem 15.10.2 looks pretty scary, but
we proved it without any algebraic manipulations at all. The key to constructing a
combinatorial proof is choosing the set S properly, which can be tricky. Generally,
the simpler side of the equation should
provide some guidance. For example, the
3n
right side of Theorem 15.10.2 is n , which suggests that it will be helpful to
choose S to be all n-element subsets of some 3n-element set.
15.11
References
[6], [10], [17]
Problems for Section 15.2
Practice Problems
Problem 15.1.
Alice is thinking of a number between 1 and 1000.
What is the least number of yes/no questions you could ask her and be guaranteed
to discover what it is? (Alice always answers truthfully.)
(a)
“mcs” — 2017/6/5 — 19:42 — page 654 — #662
654
Chapter 15
Cardinality Rules
Problem 15.2.
In how many different ways is it possible to answer the next chapter’s practice
problems if:
the first problem has four true/false questions,
the second problem requires choosing one of four alternatives, and
the answer to the third problem is an integer 15 and 20?
Problem 15.3.
How many total functions are there from set A to set B if jAj D 3 and jBj D 7?
Problem 15.4.
Let X be the six element set fx1 ; x2 ; x3 ; x4 ; x5 ; x6 g.
(a) How many subsets of X contain x1 ?
(b) How many subsets of X contain x2 and x3 but do not contain x6 ?
Class Problems
Problem 15.5.
A license plate consists of either:
3 letters followed by 3 digits (standard plate)
5 letters (vanity plate)
2 characters—letters or numbers (big shot plate)
Let L be the set of all possible license plates.
(a) Express L in terms of
A D fA; B; C; : : : ; Zg
D D f0; 1; 2; : : : ; 9g
using unions ([) and set products ().
(b) Compute jLj, the number of different license plates, using the sum and product
rules.
“mcs” — 2017/6/5 — 19:42 — page 655 — #663
15.11. References
655
Problem 15.6. (a) How many of the billion numbers in the integer interval Œ1::109 
contain the digit 1? (Hint: How many don’t?)
(b) There are 20 books arranged in a row on a shelf. Describe a bijection between
ways of choosing 6 of these books so that no two adjacent books are selected, and
15-bit strings with exactly 6 ones.
Problem 15.7.
(a) Let Sn;k be the possible nonnegative integer solutions to the inequality
x1 C x2 C C xk n:
(15.11)
That is
Sn;k WWD f.x1 ; x2 ; : : : ; xk / 2 Nk j (15.11) is trueg:
Describe a bijection between Sn;k and the set of binary strings with n zeroes and k
ones.
(b) Let Ln;k be the length k weakly increasing sequences of nonnegative integers
n. That is
Ln;k WWD f.y1 ; y2 ; : : : ; yk / 2 Nk j y1 y2 yk ng:
Describe a bijection between Ln;k and Sn;k .
Problem 15.8.
An n-vertex numbered tree is a tree whose vertex set is Œ1::n for some n > 2. We
define the code of the numbered tree to be a sequence of n 2 integers in Œ1::n
obtained by the following recursive process:6
If there are more than two vertices left, write down the father of the largest leaf,
delete this leaf, and continue this process on the resulting smaller tree. If there
are only two vertices left, then stop—the code is complete.
For example, the codes of a couple of numbered trees are shown in the Figure 15.7.
6 The
necessarily unique node adjacent to a leaf is called its father.
“mcs” — 2017/6/5 — 19:42 — page 656 — #664
656
Chapter 15
Cardinality Rules
tree
1
5
2
4
6
3
1
code
65622
7
2
3
4
5
432
Figure 15.7
(a) Describe a procedure for reconstructing a numbered tree from its code.
(b) Conclude there is a bijection between the n-vertex numbered trees and sequences .n 2/ integers in Œ1::n. State how many n-vertex numbered trees there
are.
Problem 15.9.
Let X and Y be finite sets.
(a) How many binary relations from X to Y are there?
(b) Define a bijection between the set ŒX ! Y  of all total functions from X to
Y and the set Y jXj . (Recall Y n is the Cartesian product of Y with itself n times.)
Based on that, what is j ŒX ! Y  j?
(c) Using the previous part, how many functions, not necessarily total, are there
from X to Y ? How does the fraction of functions vs. total functions grow as the
size of X grows? Is it O.1/, O.jXj/, O.2jXj /,. . . ?
(d) Show a bijection between the powerset pow.X / and the set ŒX ! f0; 1g of
0-1-valued total functions on X .
(e) Let X be a set of size n and BX be the set of all bijections from X to X.
“mcs” — 2017/6/5 — 19:42 — page 657 — #665
15.11. References
657
Describe a bijection from BX to the set of permutations of X .7 This implies that
there are how many bijections from X to X?
Problems for Section 15.4
Class Problems
Problem 15.10.
Use induction to prove that there are 2n subsets of an n-element set (Theorem 4.5.5).
Homework Problems
Problem 15.11.
Fermat’s Little Theorem 9.10.88 asserts that
ap a
.mod p/
(15.12)
for all primes p and nonnegative integers a. This is immediate for a D 0; 1 so we
assume that a 2.
This problem offers a proof of (15.12) by counting strings over a fixed alphabet
with a characters.
(a) How many length-k strings are there over an a-character alphabet?
How
many of these are strings use more than one character?
Let z be a length-k string. The length-n rotation of z is the string yx, where
z D xy and the length jxj of x is remainder.n; k/.
(b) Verify that if u is a length-n rotation of z, and v is a length-m rotation of u,
then v is a length-(n C m) rotation of z.
(c) Let be the “is a rotation of” relation on strings. That is,
vz
IFF
v is a length-n rotation of z
for some n 2 N. Prove that is an equivalence relation.
7A
sequence in which all the elements of a set X appear exactly once is called a permutation of
X (see Section 15.3.3).
8 This Theorem is usually stated as
ap
1
1
.mod p/;
for all primes p and integers a not divisible by p. This follows immediately from (15.12) by canceling
a.
“mcs” — 2017/6/5 — 19:42 — page 658 — #666
658
Chapter 15
Cardinality Rules
(d) Prove that if xy D yx then x and y each consist of repetitions of some string
u. That is, if xy D yx, then x; y 2 u for some string u.
Hint: By induction on the length jxyj of xy.
(e) Conclude that if p is prime and z is a length-p string containing at least two
different characters, then z is equivalent under to exactly p strings (counting
itself).
(f) Conclude from parts (a) and (e) that p j .ap a), which proves Fermat’s Little
Theorem (15.12).
Problems for Section 15.5
Practice Problems
Problem 15.12.
Eight students—Anna, Brian, Caine,. . . —are to be seated around a circular table
in a circular room. Two seatings are regarded as defining the same arrangement if
each student has the same student on his or her right in both seatings: it does not
matter which way they face. We’ll be interested in counting how many arrangements there are of these 8 students, given some restrictions.
(a) As a start, how many different arrangements of these 8 students around the
table are there without any restrictions?
(b) How many arrangements of these 8 students are there with Anna sitting next
to Brian?
(c) How many arrangements are there with if Brian sitting next to both Anna AND
Caine?
(d) How many arrangements are there with Brian sitting next to Anna OR Caine?
Problem 15.13.
How many different ways are there to select three dozen colored roses if red, yellow,
pink, white, purple and orange roses are available?
Problem 15.14.
“mcs” — 2017/6/5 — 19:42 — page 659 — #667
15.11. References
659
Suppose n books are lined up on a shelf. The number of selections of m of the
books so that selected books are separated by at least three unselected books is the
same as the number of all length k binary strings with exactly m ones.
(a) What is the value of k?
(b) Describe a bijection between between the set of all length k binary strings with
exactly m ones and such book selections.
Problem 15.15.
Six women and nine men are on the faculty of a school’s EECS department. The
individuals are distinguishable. How many ways are there to select a committee of
5 members if at least 1 woman must be on the committee?
Class Problems
Problem 15.16.
Your class tutorial has 12 students, who are supposed to break up into 4 groups of
3 students each. Your Teaching Assistant (TA) has observed that the students waste
too much time trying to form balanced groups, so he decided to pre-assign students
to groups and email the group assignments to his students.
(a) Your TA has a list of the 12 students in front of him, so he divides the list into
consecutive groups of 3. For example, if the list is ABCDEFGHIJKL, the TA would
define a sequence of four groups to be .fA; B; C g; fD; E; F g; fG; H; I g; fJ; K; Lg/.
This way of forming groups defines a mapping from a list of twelve students to a
sequence of four groups. This is a k-to-1 mapping for what k?
(b) A group assignment specifies which students are in the same group, but not
any order in which the groups should be listed. If we map a sequence of 4 groups,
.fA; B; C g; fD; E; F g; fG; H; I g; fJ; K; Lg/;
into a group assignment
ffA; B; C g; fD; E; F g; fG; H; I g; fJ; K; Lgg;
this mapping is j -to-1 for what j ?
(c) How many group assignments are possible?
(d) In how many ways can 3n students be broken up into n groups of 3?
“mcs” — 2017/6/5 — 19:42 — page 660 — #668
660
Chapter 15
Cardinality Rules
Problem 15.17.
A pizza house is having a promotional sale. Their commercial reads:
We offer 9 different toppings for your pizza! Buy 3 large pizzas at
the regular price, and you can get each one with as many different
toppings as you wish, absolutely free. That’s 22; 369; 621 different
ways to choose your pizzas!
The ad writer was a former Harvard student who had evaluated the formula .29 /3 =3Š
on his calculator and gotten close to 22; 369; 621. Unfortunately, .29 /3 =3Š can’t be
an integer, so clearly something is wrong. What mistaken reasoning might have
led the ad writer to this formula? Explain how to fix the mistake and get a correct
formula.
Problem 15.18.
Answer the following quesions using the Generalized Product Rule.
(a) Next week, I’m going to get really fit! On day 1, I’ll exercise for 5 minutes.
On each subsequent day, I’ll exercise 0, 1, 2, or 3 minutes more than the previous
day. For example, the number of minutes that I exercise on the seven days of next
week might be 5, 6, 9, 9, 9, 11, 12. How many such sequences are possible?
(b) An r-permutation of a set is a sequence of r distinct elements of that set. For
example, here are all the 2-permutations of fa; b; c; d g:
.a; b/
.b; a/
.c; a/
.d; a/
.a; c/
.b; c/
.c; b/
.d; b/
.a; d /
.b; d /
.c; d /
.d; c/
How many r-permutations of an n-element set are there? Express your answer
using factorial notation.
(c) How many nn matrices are there with distinct entries drawn from f1; : : : ; pg,
where p n2 ?
Problem 15.19. (a) There are 30 books arranged in a row on a shelf. In how many
ways can eight of these books be selected so that there are at least two unselected
books between any two selected books?
“mcs” — 2017/6/5 — 19:42 — page 661 — #669
15.11. References
661
(b) How many nonnegative integer solutions are there for the following equality?
x1 C x2 C C xm D k:
(15.13)
(c) How many nonnegative integer solutions are there for the following inequality?
x1 C x2 C C xm k:
(15.14)
(d) How many length m weakly increasing sequences of nonnegative integers k
are there?
Homework Problems
Problem 15.20.
This problem is about binary relations on the set of integers in the interval Œ1::n
and digraphs whose vertex set is Œ1::n.
(a) How many digraphs are there?
(b) How many simple graphs are there?
(c) How many asymmetric binary relations are there?
(d) How many linear strict partial orders are there?
Problem 15.21.
Answer the following questions with a number or a simple formula involving factorials and binomial coefficients. Briefly explain your answers.
(a) How many ways are there to order the 26 letters of the alphabet so that no two
of the vowels a, e, i, o, u appear consecutively and the last letter in the ordering
is not a vowel?
Hint: Every vowel appears to the left of a consonant.
(b) How many ways are there to order the 26 letters of the alphabet so that there
are at least two consonants immediately following each vowel?
(c) In how many different ways can 2n students be paired up?
(d) Two n-digit sequences of digits 0,1,. . . ,9 are said to be of the same type if the
digits of one are a permutation of the digits of the other. For n D 8, for example,
the sequences 03088929 and 00238899 are the same type. How many types of
n-digit sequences are there?
“mcs” — 2017/6/5 — 19:42 — page 662 — #670
662
Chapter 15
Cardinality Rules
Problem 15.22.
In a standard 52-card deck, each card has one of thirteen ranks in the set R and one
of four suits in the set S where
R WWD fA; 2; : : : ; 10; J; Q; Kg;
S WWD f|; }; ~; g:
A 5-card hand is a set of five distinct cards from the deck.
For each part describe a bijection between a set that can easily be counted using
the Product and Sum Rules of Ch. 15.1, and the set of hands matching the specification. Give bijections, not numerical answers.
For instance, consider the set of 5-card hands containing all 4 suits. Each such
hand must have 2 cards of one suit. We can describe a bijection between such hands
and the set S R2 R3 where R2 is the set of two-element subsets of R. Namely,
an element
.s; fr1 ; r2 g; .r3 ; r4 ; r5 // 2 S R2 R3
indicates
1. the repeated suit s 2 S,
2. the set fr1 ; r2 g 2 R2 of ranks of the cards of suit s and
3. the ranks .r3 ; r4 ; r5 / of the remaining three cards, listed in increasing suit
order where
| } ~ :
For example,
.|; f10; Ag; .J; J; 2//
! fA|; 10|; J }; J ~; 2g:
(a) A single pair of the same rank (no 3-of-a-kind, 4-of-a-kind, or second pair).
(b) Three or more aces.
Problem 15.23.
Suppose you have seven dice—each a different color of the rainbow; otherwise
the dice are standard, with faces numbered 1 to 6. A roll is a sequence specifying a value for each die in rainbow (ROYGBIV) order. For example, one roll is
.3; 1; 6; 1; 4; 5; 2/ indicating that the red die showed a 3, the orange die showed 1,
the yellow 6,. . . .
“mcs” — 2017/6/5 — 19:42 — page 663 — #671
15.11. References
663
For the problems below, describe a bijection between the specified set of rolls
and another set that is easily counted using the Product, Generalized Product, and
similar rules. Then write a simple arithmetic formula, possibly involving factorials
and binomial coefficients, for the size of the set of rolls. You do not need to prove
that the correspondence between sets you describe is a bijection, and you do not
need to simplify the expression you come up with.
For example, let A be the set of rolls where 4 dice come up showing the same
number, and the other 3 dice also come up the same, but with a different number.
Let R be the set of seven rainbow colors and S WWD Œ1::6 be the set of dice values.
Define B WWD PS;2 R3 , where PS;2 is the set of 2-permutations of S and R3
is the set of size-3 subsets of R. Then define a bijection from A to B by mapping
a roll in A to the sequence in B whose first element is a pair consisting of the
number that came up three times followed by the number that came up four times,
and whose second element is the set of colors of the three matching dice.
For example, the roll
.4; 4; 2; 2; 4; 2; 4/ 2 A
maps to
..2; 4/; fyellow,green,indigog/ 2 B:
Now by the B